AI

Article URL: https://www.anthropic.com/research/claude-scorpion-toxin-binder-design Comments URL: https://news.ycombinator.com/item?id=50170021 Points: 51 # Comments: 22

A single query is all that's required to find thousands of conversations listed by title.

arXiv:2610.26935v1 Announce Type: new Abstract: AI control evaluations often reset agents' memory between tasks, although deployed agents may retain records of prior interventions. We ask whether learning that an earlier attack was caught changes an agent's ability to evade defer-to-trusted (DtT), a protocol that replaces suspicious work with a trusted model's solution. Herein, we study cross-episode adaptation: whether intervention history changes attack behavior on a subsequent task with different objectives and a fresh environment. We evaluate this using ControlArena's Infrastructure-as-Code (IaC) and SHADE-Arena as testbeds, where agents must complete a main task while covertly pursuing a harmful side objective. For each successor task, we pair a history-conditioned attempt with the model's no-history attempt. The retained history either states that the previous submission was replaced or additionally includes the trusted model's replacement. For Claude Opus 4.8 on IaC, intervention history reduces successful sabotage. Completion of both the main and side objectives falls from 77% to 12% when the agent is told about the replacement and to 38% when it is also shown the trusted m…

arXiv:2610.27244v1 Announce Type: new Abstract: Agent systems are shifting from task-scoped tool-use agents to persistent, personalized assistants such as OpenClaw. These claw-style agents maintain configuration files (e.g., AGENTS.md), long-term memory, and reusable skills. Once poisoned, this persistent state is repeatedly reloaded into the agent's context, creating persistent threats across tasks and sessions. Existing security evaluation systems struggle to evaluate these complex claw-style systems since they rely heavily on manually constructed test cases and environments, which limits their scale and coverage; in addition, these manually designed cases quickly become outdated as agents evolve, and their evaluation settings leave a gap with how claw-style agents are deployed in practice. To address this, we introduce AutoRedClaw, an automated red-teaming framework that generates end-to-end attacks together with the environments needed to execute them. AutoRedClaw combines a strict quality-checking protocol with execution-based validation to ensure the correctness, realism, and effectiveness of generated samples, and adopts task-driven environment construction to keep costs low…

arXiv:2610.27257v1 Announce Type: new Abstract: Large language models can detect vulnerable code, yet the form of vulnerability encoding in their representations remains unclear. We analyze 8 code models spanning 6.7B-15B parameters on C programs using three datasets (DELTASECOMMITS, SVEN, and PRECISE-BUGS), comparing two diagnostic tasks on the same representations: identifying what patterns the code instantiates, and determining whether it is vulnerable. Models encode the first with high fidelity. Under the studied paired C-code datasets and linear-probing setup, vulnerability information is more accessible as a paired relational signal-whether code A is riskier than code B-than as a globally separable categorical label. This structural difference has implications for classification: the model reliably ranks which code is riskier but standard classifiers on isolated samples achieve limited accuracy because vulnerable and secure samples form overlapping distributions in representation space.

arXiv:2610.27353v1 Announce Type: new Abstract: Beneath the apparent diversity of foundation models lies an increasingly homogeneous technical stack of common training pipelines, distillation methods, and data. These shared algorithmic choices can expose an otherwise distinct model ecosystem to concentrated risks leading to a vulnerability monoculture. Yet current practice in assessing adversarial robustness is restricted to individual robustness scores and attack transfer rates, leaving open the unexplored fundamental question: how many attacks suffice for catastrophic widespread compromise? In this paper, we investigate the empirical nature of shared vulnerabilities within a model pool and introduce Systemic Adversarial Robustness (SAR), a novel measure of shared vulnerability of a model pool through the attack diversity needed to exploit it. Specifically, SAR interpolates between single-attack transferability and independent per-model robustness: small values reveal a vulnerability monoculture, whereas large values indicate that failures are dispersed across models. We theoretically formulate SAR as a partial set-cover problem with provable approximation guarantees---enabling th…

Meta changes AI scraping policy but won’t add language respecting robots.txt.

It's probably time to audit your Gemini settings.

Google has joined Amazon in releasing a developer tool that makes it easy for PyTorch users to move their apps off Nvidia GPUs.

The company says the feature is designed for situations where users might be discussing something personal or planning a surprise and want to keep those conversations private when handing their device to friends or family to show them something in another chat.

Meta argues nobody “decided” to pirate books to train AI.

Intel has long worked on tuning their oneDNN neural network library for their hardware and saw its adoption by PyTorch and other AI frameworks. AMD has maintained ZenDNN as their downstream fork that is oneDNN API compatible while catering to their Zen processor designs. As an interesting twist, AMD and Intel jointly announced today they will be collaborating around oneDNN for PyTorch and pushing the AMD Zen improvements into oneDNN...

It's a time of apocalyptic worldviews in Silicon Valley. The latest AI doom cycle is driven by effective altruism (EA), a philosophy that has lurked around the fringes for years but recently burst into the mainstream because its adherents are embedded across the AI industry, from Anthropic to newly appointed AI safety watchdog METR. EA is a utilitarian approach to charity premised on doing the most good for the most people over the longest time, and EA advocates see halting an AI-induced doomsday as critical to saving as many humans as possible. But EA has an important and equally apocalyptic rival in the crosscutting cosmologies of the tec … Read the full story at The Verge.

Spotify says it has removed more than 50 million songs considered AI-generated slop or spam from user recommendations over the past year. That's according to an announcement on Tuesday, which says Spotify has also made changes to its recommendation system to highlight more music made by human artists. "Our editors hold the same standard and don't program content identified as slop or spam," Spotify says. "Since we made these changes, we've seen a marked decrease in listener concerns about this content showing up in recommendations." Though Spotify has stopped short of mandating labels on all AI tracks, it launched an "AI Persona" badge fo … Read the full story at The Verge.

WhatsApp's help pages carried hidden instructions for its support chatbot that told it not to recommend a data protection authority or regulator to users who could no longer delete their accounts, even though WhatsApp's own privacy policy for Europe tells users they can complain to one. The text sat in the pages' code, never shown on screen, until it was removed by October 18. Whether the