AI
8 items · 19 articles in this topic
Google's Gemini 4 Hydrogen launch drove product news, while Google also faced backlash over a Gemini school pilot. Security and policy framing of AI continued, with OpenAI selling security products and analysis of Chinese models.
- 1
Google rolled out Gemini 4 Hydrogen as its default model, replacing Gemini 3.8 Flash, without the staged security-focused rollout used for earlier models. It powers AI Mode in Search and a new orchestrator in the Antigravity developer platform.
- 2
Families are pressing Google for moratoriums on AI in schools after a teen's lawsuit called the Gemini chatbot a 'suicide coach'.
- 3
A forensic analysis claims Chinese frontier model checkpoints carry fingerprints of export-controlled GPUs. This is a single-source claim from a non-mainstream site.
- 4
OpenAI launched security products pitched to organizations, while nonprofits sued AI companies citing AI-driven cyberattacks.
- 5
Alibaba's open-weight Qwen3.8 is reported to answer sensitive questions differently, reflecting a more Western worldview.
- 6
Bruce Schneier's essay, adapted from his DEF CON 34 keynote, covers AI hacking systems and AI acting autonomously.
- 7
Anthropic's status page reported elevated errors on claude.ai, Claude Code and platform.claude.com.
NotableHacker News (front page) ↗
- 8
A claim that AI agents discovered an unknown Maya site was false, because the site was already published in an academic journal last year.
More in AI · 11 articles›
Google announces Gemini 4 Hydrogen and an orchestrator in its Antigravity developer platform.
Hacker News (front page)geminigooglellmFeed excerpt
Article URL: https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/ Comments URL: https://news.ycombinator.com/item?id=50472260 Points: 140 # Comments: 26
SideTaskBench introduces a benchmark measuring whether coding agents deliver unrequested side-task artifacts while performing authorized repository repairs.
arXiv cs.CR (Cryptography & Security)coding-agentsbenchmarkllmFeed excerpt
arXiv:2611.14548v1 Announce Type: new Abstract: Can a coding agent fix a user's repository while also completing and delivering work the user never requested? In this paper, we introduce SideTaskBench to measure this failure where third-party technical references induce agents to produce independent side-task artifacts during an authorized repair. Repair success alone misses this behavior. Our key measurement contribution is to pair each repair with a self-contained side task and verify side tasks' delivery success rate along with the repair success rate. The benchmark contains 25 Python repairs and 190 side-task units. Its four-level similarity index defines 300 executable repair--side-task pairs. The experiments show roughly 40% valid delivery under wrapped requests for two models, alongside repair success close to their clean baselines. The completed v10 similarity run uses DeepSeek-V4-Flash on all 300 pairs, with eight repetitions per pair (2,400 runs). Conditional delivery ranges from 20.48% for L0 unrelated tasks to 57.44% for L3 mechanism-overlapping tasks, while the L1-L0 difference is 5.37 percentage points and remains sensitive to task composition. The central insight is …
Researchers test sandbagging detectors against adaptive attacks by fine-tuning small open-weight models to underperform on GSM8K when an audit marker appears.
arXiv cs.CR (Cryptography & Security)ai-safetysandbaggingevaluationFeed excerpt
arXiv:2611.14950v1 Announce Type: new Abstract: Developers, auditors and researchers read evaluation scores as evidence of what a model can do. A model that sandbags, scoring low on purpose while keeping the capability, turns a low score into false reassurance. Sandbagging detectors are usually tested once against a fixed attack, but a real attacker sees what gets flagged and adapts. We build sandbagging models by fine-tuning small open-weight models to fail math word problems (GSM8K) they can otherwise solve, but only when an audit marker is in the prompt. On Qwen 2.5-3B, we adapt the attack over two retraining rounds against a text-based audit. Two rounds of retraining cut the text audit's AUC from 0.978 to 0.560 while the marker still removes 40-49 points of accuracy. The attack worked in 7 of 8 retrained seeds, and the audit missed every retraining (AUC 0.451-0.576), including one Llama run. Activation probes also flag 99% of benign fine-tunes' examples, and matched probes transfer to only 1 of 5 held-out training runs. Eight further trigger attacks are all active under their own prompt but not under the detectors' prompt. A detector that catches a fixed attack has not been sho…
Paper shows private information can leak from fixed-size recurrent state in latent-state LLMs during inference.
arXiv cs.CR (Cryptography & Security)llmprivacyresearchFeed excerpt
arXiv:2611.16048v1 Announce Type: new Abstract: Personalized LLM services place private facts and application instructions into a prompt, then retain representations of that context for later inference. Latent-state LLMs compress part of this history into fixed-size recurrent state. The information preserved by its learned updates is implicit, and later computation can expose it even when an observer receives neither the private text nor the responses. We investigate privacy leakage from recurrent state through partial-state probing and physical observation of its use. Our methodology separates recoverable information, persistence through public dialogue, causal influence on later decisions, and leakage through power. Experiments with released hybrid implementations show that a small state observation supports multiple predictions about private facts and response rules, and that personal attributes generalize to unseen authors. The information remains inferable after public dialogue and changes later decision margins. Complete-read power recovers private facts under a fixed execution schedule. On a partitioned GPU, an unprivileged account infers age from whole-device power during p…
T-SemAttack shows small perturbations can transferably degrade speech tokenizers, hurting ASR and large audio-language model performance.
arXiv cs.CR (Cryptography & Security)adversarialspeechllmFeed excerpt
arXiv:2611.14929v1 Announce Type: cross Abstract: Speech tokenizers serve as speech front ends, encoding input speech signals into representations for downstream models, such as automatic speech recognition (ASR) systems and large audio-language models (LALMs). In this paper, we find that speech tokenizers are fragile at the semantic level: small perturbations can disrupt their semantic encoding process and consequently cause severe degradation in downstream ASR and LALM performance. To study this vulnerability, we propose T-SemAttack, a transferable semantic attack based on a semantic encoder ensemble. By jointly drifting the representation spaces of multiple semantic encoders, T-SemAttack effectively disrupts the semantic content of speech while preserving perceptual quality, thereby inducing error features. We further analyze layer-wise representation drift and cross-model gradient alignment. Our findings show that perturbations can induce substantial shifts in deep semantic representations and that the input gradients of different surrogate encoders exhibit directional alignment. These observations suggest that shared adversarially sensitive directions in the input space may be…
Google released Gemini 4 Hydrogen, a model that better handles multi-part and complex questions across Search and the Gemini app.
The VergegeminigooglellmFeed excerpt
Google says its new Gemini 4 Hydrogen model can give more helpful answers to complex questions, including "questions with multiple parts, hard-to-understand typos, or highly specific requests." The new model can break down your questions and address each part in its responses. It's available starting today in search, the Gemini app, all tiers of Google AI Studio, the Gemini API, Gemini for business, and in Google Antigravity for Google AI Pro and Ultra subscribers. Robby Stein, VP of product at Google Search, says it's "now the new default model in AI Mode & being used for your hardest questions in AI Overviews." Google says Gemini 4 Hydrog … Read the full story at The Verge.
Apple published a developer page for its Foundation Models framework, giving apps access to on-device language models.
Hacker News (front page)appleon-device-aideveloper-toolsFeed excerpt
Article URL: https://developer.apple.com/foundation-models-framework/ Comments URL: https://news.ycombinator.com/item?id=50474361 Points: 93 # Comments: 12
Google upgrades AI Mode in Search, powered by Gemini 4 Hydrogen, to better handle complex queries with richer visual responses.
TechCrunchgooglegeminisearchllmFeed excerpt
Powered by Gemini 4 Hydrogen, Google's AI Mode is getting better at understanding complex queries, surfacing relevant links, and delivering richer visual responses.
Google announced Gemini 4 Hydrogen, a new model in its Gemini family, per a Google blog post.
Hacker News (front page)geminigooglellmFeed excerpt
Article URL: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/ Comments URL: https://news.ycombinator.com/item?id=50471431 Points: 23 # Comments: 1
United Airlines is restricting AI agents from handling check-ins because safety checks require a human.
The Registerai-agentsairlinespolicyFeed excerpt
AI agents face restrictions over safety checks that require a human touch
A GitHub project, smolmaxxing, aims to run LLMs on small and older devices.
Hacker News (front page)llmopen-sourceedge-aiFeed excerpt
Article URL: https://github.com/hyperionaidev/smolmaxxing/ Comments URL: https://news.ycombinator.com/item?id=50468888 Points: 106 # Comments: 41