Cyber security
8 items · 58 articles in this topic
Remote code execution in widely deployed server software dominated the day, with several flaws already exploited. Supply-chain and service-provider breaches continued alongside, and a long tail of agent-security research papers appeared.
- 1
Apache Roller CVE-2026-129448, a CVSS 9.8 unauthenticated deserialization flaw, is being exploited in the wild.
- 2
F5 disclosed six critical BIG-IP vulnerabilities that let unauthenticated attackers crash the TMM traffic process and possibly achieve RCE in specific configurations.
- 3
CISA added four exploited flaws to KEV: Synology DSM, BerriAI LiteLLM, Apache Axis and Vanderbilt REDCap. They are linked to ransomware, cryptomining and LLMjacking attacks.
- 4
Ruby fixed seven memory-safety flaws in its regex and string code, five with proof-of-concept RCE. Verialabs separately disclosed Ruby2Shell, a pre-authentication RCE on stock Rails.
- 5
A researcher published proof-of-concept exploits that give low-privilege database accounts code execution on PostgreSQL, MySQL, ClickHouse and Oracle.
- 6
Terra Security published details of CVE-2026-99312, a KVM guest-to-host escape in Linux virtualization. The summary gives no patch details.
- 7
Attackers breached an N-able server in its Mail Assure and Spam Experts cloud, disabling management interfaces and APIs across three regions.
- 8
An October attack on service provider Wellington IT exposed payee data at 40 Irish credit unions, with notification delayed until November. Separately, TfL took its Oyster photocard site offline after customer accounts were compromised, potentially affecting about 265,000 customers.
More in Cyber security · 50 articles›
Attackers are exploiting PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619 against internet-exposed Palo Alto Networks PAN-OS management interfaces for unauthenticated code execution.
The Hacker Newscveexploitedpalo-altofirewallFeed excerpt
One of the early targets of PHP2Shell exploitation was the firewall. On October 28, Palo Alto Networks confirmed that attackers were exploiting the two PHP vulnerabilities, CVE-2026-109618 and CVE-2026-109619, against PAN-OS management interfaces exposed to the internet. Both flaws let anyone execute code on the web server without authentication. The disclosure itself went a little wrong.
F5 disclosed CVE-2026-123007 (nginx2root), a decade-old NGINX HTTP/2 flaw allowing unauthenticated remote code execution, fixed in versions 1.31.9 and 1.30.7.
The Hacker NewscvenginxrcepatchFeed excerpt
A flaw in NGINX's HTTP/2 support can allow an attacker to run code on servers with HTTP/2 enabled, without needing to log in. The decade-old flaw, CVE-2026-123007, dubbed "nginx2root," has been present in NGINX's open-source code since version 1.9.5, released on September 22, 2015. Fixed versions 1.31.9 and 1.30.7 came out today. F5, the company behind NGINX, says attackers can run code where
Xint disclosed five NGINX vulnerabilities, including an unauthenticated RCE over HTTP/2, plus an SCTP kernel LPE with no patch yet.
LWN.netnginxrcekernelvulnerabilityFeed excerpt
Xint is reporting that it has discovered a collection of five critical vulnerabilities in the NGINX web server. The worst flaw, it says, would allow an unauthenticated attacker to execute commands on an affected nginx server with HTTP/2 enabled. The company also reports a local-privilege escalation (LPE) kernel vulnerability, which was introduced in the 4.15 kernel, in the Stream Control Transmission Protocol (SCTP) subsystem. If chained with the NGINX vulnerability, it would allow an attacker to gain root privileges on the host. A kernel patch for this is not yet available. Xint recommends updating to NGINX 1.31.9, which was just released, and to block the SCTP module on hosts that do not need it until a kernel fix is available. See the mitigations section of their report for more information.
Wiz details pre-authentication remote code execution vulnerabilities in NGINX's HTTP/2 handling affecting web servers.
Hacker News (front page)cvenginxrceweb-serverFeed excerpt
Article URL: https://www.wiz.io/blog/nginx-http2-rce-vulnerabilities Comments URL: https://news.ycombinator.com/item?id=50547374 Points: 400 # Comments: 132
Cloudflare released six WAF managed rules against the critical Java4Shell flaw CVE-2026-114938 in OpenJDK's Base64 decoder.
Cloudflare BlogcvejavarcewafFeed excerpt
Java4Shell is a critical memory-corruption flaw in OpenJDK's Base64 decoder that can lead to remote code execution. Cloudflare has released six WAF managed rules for paid plans that logs matching requests by default. Patch your Java runtimes now.
The Drupal Security Team patched a critical remote code execution flaw, and warns that exploits may appear quickly.
BleepingComputercvedrupalrceFeed excerpt
The Drupal Security Team has patched a critical remote code execution vulnerability that can let attackers take over vulnerable sites. [...]
Transport for London took the Oyster photocard website offline after attackers compromised customer accounts, potentially affecting about 265,000 customers.
The Registerbreachtflcustomer-dataFeed excerpt
Around 265,000 customers potentially affected as London transport operator grapples with another security breach
NVIDIA is investigating claims that the Gentlemen ransomware group breached it after listing the company on its leak site.
BleepingComputerransomwaredata-breachnvidiaFeed excerpt
NVIDIA is investigating claims of a security breach after the Gentlemen ransomware group has added the company to its dark web data leak site. [...]
Verialabs disclosed a pre-authentication remote code execution flaw affecting stock Ruby on Rails installations.
Hacker News (front page)rcerailsrubyvulnerabilityFeed excerpt
Article URL: https://verialabs.com/ruby2shell/ Comments URL: https://news.ycombinator.com/item?id=50556287 Points: 51 # Comments: 11
The PHP maintainers announced a November 19 security release fixing critical flaws that affect all supported PHP versions.
The Hacker NewsphpadvisorypatchFeed excerpt
The maintainers of the PHP programming language on Wednesday issued an advisory announcing an upcoming security release that addresses critical flaws. The release is scheduled for November 19, 2026, between 2 and 2:30 p.m. UTC. "These releases are part of PHP's regular four-week release cycle and will address security vulnerabilities affecting all supported PHP versions," the maintainers said.
Ukrainian authorities and the FBI arrested four suspected members of a ransomware group tied to over $40.5M in ransoms.
BleepingComputerransomwarearrestlaw-enforcementFeed excerpt
Ukrainian authorities, working with the FBI, have arrested four suspected members of an international ransomware organization linked to more than $40.5 million in ransom payments from at least 28 companies worldwide. [...]
PHP is preparing a security release that will fix at least one dangerous remote code execution bug.
The RegisterphprcepatchFeed excerpt
At least one dangerous RCE bug will get a fix on Thursday
AI-assisted analysis found more critical bugs in NGINX, an open source web server running on about 225 million websites.
The Registernginxvulnerabilityopen-sourceaiFeed excerpt
Open source NGINX joins patchapalooza
SANS discusses recent critical Java and NGINX vulnerabilities and a Microsoft 365 outage in weighing its InfoCon threat level.
SANS Internet Storm CenterjavanginxinfoconFeed excerpt
About a month ago, we went to Infocon Yellow for the first time in a few years. We decided to address the php2shell vulnerability and alert our readers to its significant impact. Following php2shell, we had a number of vulnerabilities and events that may also have scored high enough to trigger an Infocon "Yellow" [infocon]. In the last few days, we had a major Microsoft 365 outage, a critical vulnerability in Java, and a critical vulnerability in NGINX, which was disclosed today.
A critical PHP security release is upcoming on 18 November 2026, according to a Lobsters notice.
LobstersphppatchadvisoryFeed excerpt
Comments
A reported cyberattack on IT provider Wellington IT exposed member data from around 40 Irish credit unions.
BleepingComputerdata-breachcredit-unionsirelandFeed excerpt
A reported cyberattack on Wellington IT has exposed sensitive information belonging to members of 50 credit unions, the majority of which are located in Ireland. [...]
Attackers compromised Microsoft's official Agent Framework package on PyPI and published two malicious versions that steal cloud credentials.
BleepingComputersupply-chainpypicredentialsmalwareFeed excerpt
Hackers compromised Microsoft's official Agent Framework Python package on PyPI, publishing two malicious versions designed to steal cloud credentials and other sensitive information from developers' systems. [...]
Krebs reports attackers abused the Optimizely A/B testing platform to serve malware to visitors of major websites.
Hacker News (front page)malwaresupply-chainwebFeed excerpt
Article URL: https://krebsonsecurity.com/2026/11/attackers-used-optimizely-to-a-b-test-malware-on-top-websites/ Comments URL: https://news.ycombinator.com/item?id=50547236 Points: 18 # Comments: 1
Bitdefender reports a suspected Chinese APT has rebuilt the EggStreme malware to evade detection while targeting Philippine government organizations.
The Hacker NewsmalwareaptespionageFeed excerpt
The suspected Chinese advanced persistent threat (APT) group behind the sophisticated EggStreme malware has rebuilt its toolkit as it continues to target government organizations in the Philippines. "When attackers are exposed, the cheap response is to rotate," Bitdefender said in a technical report shared with The Hacker News. "A file hash changes with every recompile, a domain costs a few
A GitHub pull request to the OpenJDK repository addresses a bug nicknamed java4shell that was reported in September 2026.
Hacker News (front page)javavulnerabilityFeed excerpt
Article URL: https://github.com/openjdk/jdk/pull/32781 Comments URL: https://news.ycombinator.com/item?id=50558906 Points: 15 # Comments: 1
Microsoft is testing an Edge Lockdown Mode that disables WebGL, WebGL2, and WebGPU APIs to reduce attack surface.
BleepingComputerbrowsermicrosofthardeningFeed excerpt
Microsoft is developing a new Edge security feature to reduce the attack surface by disabling the WebGL, WebGL2, and WebGPU APIs. [...]
A GitHub security advisory GHSA-34m4-m9fh-gqp5 covers a vulnerability named Ruby2Shell in Ruby software.
LobstersadvisoryrubyghsaFeed excerpt
Comments
Proposes state-transfer auditing and controls for long-horizon LLM agents to catch prompt injections and unverified memory writes.
arXiv cs.CR (Cryptography & Security)llm-agentsprompt-injectionauditingFeed excerpt
arXiv:2611.23346v1 Announce Type: new Abstract: Long-horizon LLM agents across codebases, web browsers, and enterprise infrastructure rely on an execution harness to move state between the working context, session history, persistent memory, and external environment. While conventional execution traces log model prompts and tool calls, critical security and reliability failures occur during state transfers: a transfer can admit unchecked content (indirect prompt injections, unverified memory writes), rewrite state lossily (context compaction dropping user constraints, memory consolidation stripping source authority), or restore state incompletely (workspace rollbacks leaving aborted KV state active on the inference server). Because current traces omit what crosses these boundaries, failure attribution and runtime monitors fail on multi-step trajectories. We present a state-transfer audit architecture for the agent harness. At each transfer boundary, we define an audit point specifying the minimal hash-chained record the harness emits, separating low-overhead deterministic checks from model-based semantic checks and binding internal context, memory, and inference-server cache state …
Presents ACaM, a method to model what agentic AI tools can access, exfiltrate, and propagate, independent of their descriptions.
arXiv cs.CR (Cryptography & Security)agentic-aicapabilitiesaccess-controlFeed excerpt
arXiv:2611.23507v1 Announce Type: new Abstract: In agentic AI systems, an LLM invokes tools that hold standing access to credentials, private data, filesystems, and external services. Such agents are increasingly deployed without an answer to a basic question: what can an agent access, exfiltrate, and propagate, independent of what it is designed for or is observed to do? The question goes unanswered because existing approaches trust a tool's natural-language description as a proxy for its behavior. But nothing enforces that the description and its implementation agree. This gap is known as Description-Code Inconsistency (DCI). DCI enables a threat class called capability laundering, in which a tool's implementation contains dangerous behavior that its description omits. The behavior may come from incidental drift or a supply-chain compromise. It runs silently whenever an LLM invokes the tool for its stated, benign purpose. We present ACaM, a neurosymbolic approach to capability modeling that derives each tool's capabilities from its code, so a laundered capability cannot hide behind a benign description. In ACaM, static analysis grounds an LLM in code-derived facts, and no LLM out…
Describes A-Tag, a cryptographic scheme binding files to verified live human presence to prove human involvement.
arXiv cs.CR (Cryptography & Security)cryptographyprovenancehuman-in-the-loopFeed excerpt
arXiv:2611.23602v1 Announce Type: new Abstract: The increasing autonomy of software agents and the proliferation of AI-generated content have weakened the reliability of digital actions as evidence of human involvement. Existing mechanisms, such as digital signatures and authentication protocols, verify identity and integrity, but do not establish whether a live human was present at the time an action was performed. In this paper, we introduce A-Tag, a system for verifiable human-in-the-loop tagging of digital artifacts. A-Tag cryptographically binds a file to a verified instance of live human presence at the moment of tagging. The resulting tag serves as a persistent and transferable proof that the action was triggered by a live human, without making claims about the origin, identity of the actor, or semantic content of the file. A-Tag is a modular framework combining liveness verification based on micro-motion signals shaped by human motor control, a secure cryptographic binding between the verified event and the artifact, and a publicly verifiable logging mechanism enabling independent validation. We formalize security properties of human presence signals, define an adversarial …
Protecting Web Applications using a Multimodal AI-based System for Classifying Web Injection Attacks3
Proposes a multimodal AI system that classifies web injection attacks from encrypted HTTPS traffic for security analysts.
arXiv cs.CR (Cryptography & Security)web-securityinjectionidsFeed excerpt
arXiv:2611.23660v1 Announce Type: new Abstract: Web applications have been essential to accessing and managing online services. Made available with weaknesses, much due to their rapid development and human error, they are exposed to frequent web injection attacks, leaving security analysts overwhelmed with threats and alerts that often lack critical details for proper risk assessment. Detecting these attacks requires inspecting the payload of network traffic packets, which becomes inaccessible when traffic is encrypted via HTTPS, as encryption hides the content of the traffic. This paper introduces a Multimodal Artificial Intelligence (AI) system to protect online services that rely on web application security by detecting and classifying web injection attacks, and provides security analysts with information to support attack prevention and incident response. The system employs unsupervised Machine Learning (ML) to flag potentially anomalous Network Flows (NetFlows). The corresponding data recorded in Web Server Logs (WebLogs) is then analysed using a heuristic framework based on Natural Language Processing (NLP) and supervised ML to detect and classify web injection attacks. Final…
FlowSH correlates anomalous network flows with endpoint logs to detect SQL injection attacks hidden by HTTPS encryption.
arXiv cs.CR (Cryptography & Security)sqliintrusion-detectionlogsFeed excerpt
arXiv:2611.23669v1 Announce Type: new Abstract: Web applications compose the core of internet services, rendering them prominent targets for audacious attackers. Among web attacks, SQL injection (SQLi) remains a persistent, damaging web applications connected to databases. The widespread adoption of HTTPS has been a critical advancement for online security, ensuring the confidentiality and integrity of communications by encrypting network payloads. However, it also renders traditional Network Intrusion Detection Systems (NIDS) less effective, as these typically rely on packet inspection of plaintext traffic. To access encrypted payloads, some NIDS resort to TLS-interception proxies, which leads to data leaks and security risks. This paper presents FlowSH, a system for detecting web attacks that is agnostic to whether the traffic is encrypted (HTTPS) or not (HTTP). Rather than decrypting packets in transit, FlowSH correlates flow-level metadata (NetFlows) with endpoint logs, which inherently contain decrypted payloads after reaching the web server. The system combines unsupervised machine learning to flag anomalous NetFlows and similarity search over server-side logs using Locality …
Compares the Jev 1.13 typed decision model with Gemini 3.5 Flash-Lite for phishing email detection accuracy, cost, and confidence.
arXiv cs.CR (Cryptography & Security)phishingemail-securityllmFeed excerpt
arXiv:2611.23683v1 Announce Type: new Abstract: Phishing remains a major source of credential theft and fraudulent payments, and email security systems must classify large message volumes reliably, quickly, and at manageable cost. Large Language Models (LLMs) offer flexible text understanding, but their token-based generation increases response time and API cost, and the reliability of their confidence values requires measurement. Jev 1.13, a typed decision model from TypeSafe AI, instead returns a structured outcome with a confidence value. This paper compares Jev 1.13 with Gemini 3.5 Flash-Lite, NVIDIA Nemotron 3 Ultra, and Meta Llama 4 Maverick on a labelled dataset of 1500 emails, covering classification performance, decision latency, token consumption, API cost, and probability calibration. Jev achieves the highest accuracy (79.67%), precision (71.43%), recall (97.97%), and F1 score (82.62%). It also records the shortest batch time (173.88 seconds), the lowest token count (597,895), the lowest cost (USD 0.015103 per 1000 decisions), and the lowest Brier score (0.1698) and ECE (0.1594). These results suggest that Jev is a promising option for fast, structured, high-volume triag…
Measures how envelope-layer defenses perform against tool-poisoning and memory-poisoning attacks on a single LLM agent.
arXiv cs.CR (Cryptography & Security)llm-agentprompt-injectionarxivFeed excerpt
arXiv:2611.23844v1 Announce Type: new Abstract: The envelope layer is the channel through which malicious content enters an LLM agent. Prior work established label and memory effects on multi-agent A2A envelopes, but the weight structure when MCP tools and memory coexist in a single agent remains unclear. Our threat model assumes an attacker pre-plants a malicious tool and poisons long-term memory; a benign user request inadvertently triggers it, and the attacker never enters the conversation. Following GDA Measurement and A/B/C/D, we measure how labels, memory, tool roles, and the backend shape invocation decisions across label-memory conflict, tool roles, robustness, and tool generalization. A tool-side forced-verification label substantially reduces semantic breach, but strong success memory pulls it back; a memory-side downgrade statement returns the outcome to the level of the label alone. The label works through mandatory implementation inspection plus default denial, not length or format; memory authorization wording erodes it with dose, while position and multi-turn pressure do not change the defense direction. Tool roles divide label effects into substitution, suppression,…
Proposes intent contracts as an authorization interface that keeps tools away from the planning model in LLM agents.
arXiv cs.CR (Cryptography & Security)llm-agentauthorizationarxivFeed excerpt
arXiv:2611.24025v1 Announce Type: new Abstract: Intent contracts are an authorization interface for tool-using LLM agents in which the planning model holds no tools. For each step it submits a short logged contract: a goal, a verbatim quotation of the user, and scope fields. Code validates the contract, and a small trusted model checks it against the user's words and writes the tool call itself. Provided the validation and dispatch code is correct, every executed call is written by the trusted model, the planner's influence on a call is confined to a contract logged before it, and the trusted side decides how much of a request to carry out. We compare contracts with two reviewers of planner-written commands: one sees only the user's words and the command, the information boundary of a reasoning-blind permission classifier, and one also sees the planner's goal and a checked quotation. With gpt-6.1-terra as planner and GLM-5.3-Flash as the trusted model on three AmPermBench task families, every object changed under contracts is named in the goal of the contract that produced the change; under review the planner names objects in its commands and rarely in its goals. When the planner a…
Proposes control-flow integrity enforcement for LLM agent skills to detect deviations and prompt-injected procedures.
arXiv cs.CR (Cryptography & Security)llm-agentprompt-injectionarxivFeed excerpt
arXiv:2611.24185v1 Announce Type: new Abstract: Control-flow integrity (CFI) confines a program to a control-flow graph fixed ahead of time. Procedural agent skills move this control flow into prompt text, and a large language model (LLM) agent decides at runtime whether each step runs in its reasoning, a tool call, or a script. The procedure's control skeleton survives, but its enforceability depends on where each step runs. Many violations, such as overwriting a file before validation, use only operations the skill requires, so permission allowlists admit them. A deviating or prompt-injected model can also evade a monitor by regime migration, for example by claiming that a check passed. We present SkillCFI, which carries CFI over to skill-based agents. Its compiler confines LLM interpretation to compile time and derives the agent's instructions and an executable contract from one skill. By a formal enforceability criterion, it enforces obligations whose evidence crosses the tool boundary and rejects a skill whose mandatory obligation leaves nothing to check. A deterministic runtime enforcer admits a tool request only on trusted evidence and refuses unmediated channels such as inl…
Evaluates which prompt injection operations survive when untrusted documents are rendered as page images for multimodal models.
arXiv cs.CR (Cryptography & Security)prompt-injectionmultimodalarxivFeed excerpt
arXiv:2611.24231v1 Announce Type: new Abstract: Multimodal language models increasingly answer questions over external documents, exposing them to indirect prompt injection: commands planted in untrusted content to hijack the response. Rendering untrusted content as page images lowers attack success on average, leaving open which injected operations survive rendering, at what utility cost, and whether failed image attacks reflect unread commands or commands read but not followed. We pair text and page-image versions of 191 English question-answering sources across six models, measuring marker appending, answer formatting, and answer replacement alongside clean Question-Answering utility. Rendering reduces append and format execution for every model, yet Qwen3-VL-8B still executes image-borne replacement at 77.96\%. Protection also depends on target type, rendering settings, wording, and repetition, and a search over eight fixed attack candidates largely closes the image deficit in appending and formatting for Qwen and Gemini, though not for GPT-5-nano. Utility ratios on original questions range from 93.31\% to 101.61\%, but Qwen falls to 61.99\% when answers combine two evidence sp…
Shows that shuffled obfuscation of LLM KV caches fails under repeated tokens, enabling token recovery by infrastructure operators.
arXiv cs.CR (Cryptography & Security)kv-cachellm-servingarxivFeed excerpt
arXiv:2611.24276v1 Announce Type: new Abstract: LLM serving systems store, offload, and share KV caches, exposing token-level representations to infrastructure operators and co-tenants. Shuffled obfuscation combines secret linear transforms with fresh permutations, claiming security from b! possible shufflings. We show that input symmetry invalidates this counting argument: security depends on distinguishable outputs under chosen inputs. Repeated tokens produce identical first-layer Value rows, making every row permutation leave the data unchanged. Only auxiliary structures such as a single-row recovery marker vary, reducing the effective state space from b! to b. We construct Repeat-to-Reveal. The attacker collects the b distinguishable states from its own probes and exploits their rank-one differences to recover an equivalent denoiser without knowing the defender's secret parameters. A one-time calibration transfers across requests sharing the same secret epoch, enabling passive token-multiset recovery from subsequently observed cache blocks through public-model dictionary matching. We instantiate the attack on official KV-Cloak and evaluate 3,000 real WildChat user prompts acros…
AutoPI proposes a taxonomy-guided adaptive red-teaming method to test prompt injection attacks against LLM agents more broadly.
arXiv cs.CR (Cryptography & Security)prompt-injectionllm-agentsred-teamingFeed excerpt
arXiv:2611.24313v1 Announce Type: new Abstract: Prompt injection is the dominant unsolved security risk for LLM agents, and the community now agrees that robustness must be measured against adaptive attackers rather than fixed benchmark injections. Many adaptive red teamers exist, and most are built to maximize attack success rate (ASR). Yet ASR counts how often a target breaks, not in how many different ways. A red teamer can achieve a high ASR by reusing a few techniques that work, leaving the others untested and absent from the attacks used to harden models. We present AutoPI, an adaptive red teamer designed to find many different ways to break a target, not just to break it often. AutoPI uses PIAtlas, a taxonomy of prompt-injection techniques, both to guide its search and to measure coverage. For each scenario, an attacker LLM writes an injection for every sub-technique in the taxonomy and then reworks the ones that fail using feedback from an LLM judge. The method is black-box and training-free, it plugs into existing benchmarks in place of their fixed attack templates, and it stays up to date as the taxonomy evolves. We evaluate AutoPI on AgentDojo and IPI Arena against open-…
PIPETTE uses purple-teaming with reinforcement learning to generate indirect prompt injection attacks and harden LLM agents via adversarial training.
arXiv cs.CR (Cryptography & Security)prompt-injectionllm-agentsadversarial-trainingFeed excerpt
arXiv:2611.24362v1 Announce Type: new Abstract: Large language model (LLM) agents routinely interact with untrusted external data, exposing them to indirect prompt injection (IPI) attacks that hijack agents into executing attacker-chosen tasks. To uncover these vulnerabilities at scale, automated red-teaming methods often leverage reinforcement learning (RL) to generate attacks that achieve high attack success rates (ASR). Model developers use the attacks discovered during internal red-teaming to harden agents through adversarial training before deployment, a collaborative offensive-defensive process known in cybersecurity as purple-teaming. However, we show that achieving high ASR during red-teaming does not guarantee robustness gains from subsequent adversarial training. Specifically, we find that two state-of-the-art RL-based red-teaming methods, RL-Hammer and PISmith, suffer from mode collapse, repeatedly discovering similar attack strategies whose additional successes add no additional robustness improvements to adversarial training. To address this gap, we introduce PIPETTE, a unified prompt injection purple-teaming framework that effectively converts red-teaming discoveries …
Researchers show adversaries renting data center workstations could synchronize unoptimized LLM training to create large periodic power oscillations, and propose mitigations.
arXiv cs.CR (Cryptography & Security)data-centerattackllmFeed excerpt
arXiv:2611.23666v1 Announce Type: cross Abstract: Multi-tenant data center (DC) facilities rent computing capacity on demand, placing controllable load within reach of anyone willing to pay for it. This paper considers a stealthy adversary who rents workstations and runs deliberately unoptimized large language model (LLM) training, synchronizing their consumption into large periodic oscillations while each workload remains individually indistinguishable from benign training. The periodicity of such workloads is characterized experimentally and used to model facility consumption, showing that oscillations emerge from the interaction of concurrent workloads, whether aligned by an adversary or arising naturally. A combined small- and large-signal assessment, under a hardware-grounded threat model whose feasibility is supported by experiments on a supercomputer, locates the oscillatory weak points of the network. MW-scale oscillations are shown to saturate the generators' control limiters, amplifying their effect, while Floquet analysis reveals a lightly damped mode emerging near the forcing frequency. To counter both origins, signal-energy-based detection and mitigation are proposed, …
DetectViT: Test-time Backdoor Detection for Vision Transformers via Inter-Head Attention Discrepancy3
DetectViT proposes a test-time method detecting backdoor-triggered inputs in Vision Transformers by analyzing inter-head attention discrepancies.
arXiv cs.CR (Cryptography & Security)backdoorvision-transformerml-securityFeed excerpt
arXiv:2611.23770v1 Announce Type: cross Abstract: Vision Transformers (ViTs) now underpin the visual encoders of most multimodal models, but this success comes at a cost: relying on third-party pretrained checkpoints leaves these systems open to backdoor attacks. Existing test-time detection methods rely on either the predicted label or intermediate embeddings, limiting their applicability across diverse tasks and incurring substantial computational overhead. In this work, we study how backdoor triggers affect the multi-head attention mechanism of ViTs and identify two distinct anomalous patterns: $\textit{complete attention hijacking}$, causing the attention to exhibit abnormally high inter-head spatial overlap, and $\textit{partial attention hijacking}$, causing an abnormally large variance in the concentration of the attention distribution. Motivated by these findings, we propose $\textbf{DetectViT}$, a test-time backdoor detection method that requires neither training data, model outputs, nor any prior knowledge of the trigger. DetectViT quantifies the two hijacking phenomena via an inter-head consistency score and an inter-head entropy variance score, with thresholds estimated…
RECLAIM is a training-free inference-time defense that uses a vision-language-action policy's attention to purify adversarial patch attacks on robots.
arXiv cs.CR (Cryptography & Security)adversarial-patchroboticsdefenseFeed excerpt
arXiv:2611.24494v1 Announce Type: cross Abstract: Vision-language-action policies control embodied agents through a perception-action loop grounded in physical interaction, where adversarial patches can corrupt action predictions and thereby alter subsequent observations. To mitigate these threats, many existing defenses rely on policy adaptation or auxiliary models, but require additional training or inference. We introduce RECLAIM, a training-free, inference-time defense that uses a frozen VLA's own attention response to local visual interventions to guide purification during closed-loop embodied manipulation. Our key insight is that attention changes induced by local edits provide feedback for choosing where and how to purify as the robot interacts with its environment. RECLAIM combines this feedback with robot geometry to refine patch localization and select local treatments, then revalidates them as the agent's embodied actions change the observed scene. We evaluate RECLAIM against patches generated by four adversarial attack methods across multiple manipulation task suites under frozen VLA control. Under adversarial patch attacks, RECLAIM succeeds in at least one rollout on 9…
SCR is a detector that identifies backdoor-poisoned training samples by computing clean-reference residuals in learned representations.
arXiv cs.CR (Cryptography & Security)backdoordata-poisoningml-securityFeed excerpt
arXiv:2611.24558v1 Announce Type: cross Abstract: Deep neural networks are vulnerable to backdoor attacks, in which adversaries manipulate model predictions by injecting poisoned samples into training data. Detecting these samples remains challenging because legitimate semantic variation can obscure poisoning-related signals in learned representations. Consequently, class-level statistics used by existing detectors may flag benign feature deviations while overlooking poisoned samples that blend into the class distribution. In this paper, we propose Semantic Counterfactual Residualization (SCR), which uses a small set of clean reference samples to account for benign variation through sample-specific feature reconstruction. SCR identifies poisoned samples by detecting coherent deviations from these clean baselines, without requiring prior knowledge of the attacks. We evaluate SCR through experiments covering twelve dirty-label and clean-label attacks across CIFAR-10, CIFAR-100, and Tiny ImageNet. With a small clean-reference set, SCR achieves attack-averaged true positive rates above 99.5\% and false positive rates below 2\% on each dataset, outperforming eight representative detecti…
A malicious federated learning client uses an iterative gradient inversion pipeline to recover exact sentences of other clients' training text under FedAvg.
arXiv cs.CR (Cryptography & Security)federated-learninggradient-inversionprivacyFeed excerpt
arXiv:2611.24689v1 Announce Type: cross Abstract: Malicious adversaries can employ gradient inversion attacks to reverse-engineer private data using the model updates of a round in Federated Learning. While there have been developments in gradient inversion for text, exact text recovery by a malicious client has so far remained infeasible, as FedAvg entangles the gradient contributions of many clients and obscures the signal of any individual sentence. We introduce a novel attack pipeline that unwinds this entanglement progressively through an iterative loop, in which the gradient contributions of sentences recovered exactly in one round are subtracted from the aggregate, exposing the remaining sentences for recovery in the next iteration. We show, for the first time, exact sentence recovery, with 100% precision, and on aggregates containing up to $4\times$ more tokens than the model's hidden dimension, we recover up to 80.2% of client sentences exactly, a regime where existing attacks recover none exactly.
A static-analysis pipeline uses graph kernels to attribute unknown Android APKs to specific residential proxy malware families.
arXiv cs.CR (Cryptography & Security)androidproxywaremalwareFeed excerpt
arXiv:2604.27302v2 Announce Type: replace Abstract: Android residential proxy applications represent a growing class of potentially-unwanted programs (PUPs) that covertly route third-party traffic through end-user devices, enabling ad fraud, credential abuse, and evasion of geolocation controls by sophisticated threat actors. Attributing an unknown APK to a specific proxy network remains challenging due to code reuse, SDK embedding, and obfuscation across proxy families. We present a static-analysis pipeline for automated proxyware family attribution, extracting graph-structured representations (control-flow and function-call graphs) and behavioral signatures from a labeled corpus of 3,365 Android proxy apps spanning four commercial proxy networks. We evaluate Weisfeiler-Lehman graph kernel features alone and fused with binary capability vectors across multiple classifiers. Using 5-fold DEX-grouped cross-validation to prevent data leakage, SGD achieves a macro F1 of 0.985 on the expanded dataset. To support explainability, we map classifier decisions to automatically generated Yara rules, achieving per-family accuracies up to 88.45\% after filtering non-discriminative signatures. F…
A proposal turns LLM interpretability signals such as probes and sparse autoencoders into a serving-time runtime defense against jailbreaks.
arXiv cs.CR (Cryptography & Security)jailbreakllmguardrailsFeed excerpt
arXiv:2608.25697v2 Announce Type: replace Abstract: Large language models (LLMs) are served behind layered defenses such as alignment and guardrail models, yet jailbreaks still get through. Interpretability research looks inside the model, where methods such as probes and sparse autoencoders expose signals of harmful content during generation. A signal alone, however, is not a defense. To become one, the serving system must tie each signal to the right request, apply the policy as declared, and release output only when the policy allows. Today, defenses built on such signals handle these duties on their own, so every new detector needs its own integration. Yet none of these duties depends on how a signal is computed. We present Language Model Security Modules (LMSM), which adapts the design of Linux Security Modules to LLM serving and carries out these duties once in a shared mediation path. The path hosts pluggable backends that turn model internals into evidence and policies that decide what that evidence allows. Each policy declares the evidence and evaluations it needs, and LMSM provides exactly these for each request and verifies that they took place, so a failure is never mis…
LIDAR proposes behavioral fingerprinting to verify which LLM sits behind a coding agent harness, since substitution can alter security-relevant decisions.
arXiv cs.CR (Cryptography & Security)llmcoding-agentsfingerprintingFeed excerpt
arXiv:2609.28559v2 Announce Type: replace Abstract: LLMs increasingly operate through coding-agent harnesses that inspect repositories, invoke tools, and modify files. Substituting the model behind such an agent can therefore change security-relevant decisions, including whether it verifies changes or recovers safely from failures. Existing LLM fingerprints largely infer identity from direct text or token distributions. In coding agents, these signals are mediated by system instructions, controller logic, tools, and execution feedback, limiting their transfer. We present LIDAR (LLM Identification from Decisions and Actions at Runtime), an active black-box fingerprinting method for coding-agent execution. Three coding probe pairs expose post-edit verification, transient-failure recovery, and specification--test conflict resolution under controlled changes. LIDAR represents the resulting trajectories with complementary instance-level and distribution-level features and compares them with clean references using a lightweight probabilistic identifier. It requires no access to model weights, logits, or provider internals. We evaluate 24 models from eight families on OpenCode, OpenHands,…
An ISC guest diary describes using Git, pull requests, and a human review gate to manage honeypot analysis work including AI.
SANS Internet Storm Centerhoneypotthreat-researchaiFeed excerpt
[This is a Guest Diary by Kevin Truong, an ISC intern as part of the SANS.edu BACS program]
An arXiv paper presents a previous-token prediction technique for inverting LLMs to reconstruct their prompts.
Hacker News (front page)llmprompt-injectionresearchFeed excerpt
Article URL: https://arxiv.org/abs/2607.29378 Comments URL: https://news.ycombinator.com/item?id=50554908 Points: 12 # Comments: 4
The Ruby project published advisories for multiple vulnerabilities in Regexp and String case conversion.
Hacker News (front page)rubyadvisoryvulnerabilityFeed excerpt
Article URL: https://www.ruby-lang.org/en/news/2026/11/18/multiple-vulnerabilities-in-regexp-and-string/ Comments URL: https://news.ycombinator.com/item?id=50554038 Points: 101 # Comments: 22
Amidst fears of AI cyberattacks, Satya Nadella says lots of critical infrastructure is ‘unpatchable’3
Satya Nadella warned that much critical infrastructure is unpatchable amid AI-driven cyberattack fears, downplaying a recent Outlook outage.
TechCrunchcritical-infrastructureai-threatsoutlookFeed excerpt
The Microsoft CEO downplayed the recent Outlook outage while noting that AI was causing both opportunities and new challenges for securing much of the connected world.
Coinbase says it blocked $4.7 million in crypto transfers to scammers running a deepfake Elon Musk giveaway during a Tesla livestream.
BleepingComputerscamdeepfakecoinbasefraudFeed excerpt
Coinbase says it blocked $4.7 million in cryptocurrency transfers to scammers promoting a fake giveaway using an AI-generated deepfake of Elon Musk during a Tesla livestream that attracted more than 500,000 viewers. [...]
Researchers describe three scanners, Octopus, Gecko and Tortoise, that harvest leaked AI API keys from exposed sites.
Hacker News (front page)api-keysscannerscredentialsFeed excerpt
Article URL: https://tin.computer/blog/fake-ai-crawler-secrets-scanners Comments URL: https://news.ycombinator.com/item?id=50550461 Points: 13 # Comments: 1
Citi and other banks say customers who use personal finance AI agents are responsible for losses even if they get hacked.
TechCrunchai-agentsfraudbankingFeed excerpt
Citi is the latest major financial institution to put the breaks on personal finance AI agents.