AI Safety news and analysis on Web Pulse: 28065 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A participatory auditing workshop with 16 queer community members at EurIPS 2025 traced AI harms across four case studies to problem definition and data collection stages of the development pipeline, according to the arX…
All 11 commercial AI-text humanizers on a public leaderboard changed or dropped at least one checkable unit — numbers, dates, person and organization names, or quoted spans — in 35-65% of the October 2026 documents that …
A new arXiv paper (2611.00420v1) introduces Circuit Oracle, an off-the-shelf LLM with a multi-agent harness and task-specific authored skills that answers natural-language questions about a target model by reading its at…
A formal grammar-lens benchmark derived from computation theory found that GPT-5.5 and its Pro variants accepted progressively more novel test data as test-time reasoning effort rose, with acceptance rates of 16%, 40%, 4…
A preregistered study of 23 confirmatory tests found that eliciting five weighted estimates within a single response did not uniformly debias language model judgments across anchoring, hindsight, base-rate, conjunction, …
A study on DisasterBench found that majority voting leaves an already-discovered correct workflow unselected on 10-19% of tasks per candidate pool, while structural validation never removed the last correct workflow in t…
A study posted as arXiv:2611.00672v1 found that flipping the recorded decisions of near-copy past cases from DENY to APPROVE raised approval rates by roughly 30 to 55 percentage points in DeepSeek V4.1 Flash, GPT-5.6 Lun…
A new arXiv paper (2611.00726v1) constructs a budgeted evasion attack against KGW text watermarks by replacing token-event evidence through ordinary phrase rewrites, using an independent model's proposed alternatives and…
A framework called CuriousTogether, introduced in arXiv paper 2611.01301v1, found that curiosity-guided multi-agent LLM interactions improved performance over non-curious settings on seven of eight tasks spanning reasoni…
An empirical study of 54 self-modification experiments across six language models found that whether and when a model stops rewriting its own system prompt is tied to model identity rather than the optimization mode assi…
Keyword retrieval introduced untestable or measurement-disconfirmed criteria in 46.8% of GPT-4.1-mini ECG reports across 1,083 human-validated PTB-XL patient records, according to an arXiv study (2611.01409v1) evaluating…
A new arXiv paper (arXiv:2611.01427v1) proposes a discrete-class text provenance detection framework that separates surface-level AI polishing from AI-driven changes to discourse structure, aiming to give finer-grained s…
A new arXiv paper (2611.01378v1) introduces an information-theoretic framework for designing sparse autoencoder (SAE)-based activation steering in large language models, using Jensen-Shannon divergence (JSD) between empi…
Binance said on Tuesday that a "sophisticated, multi-stage campaign" combining fake recruitment interviews, a trojanized Homebrew installer and abuse of an internal AI tool led to the theft of $6.43 billion in assets fro…
SemiAnalysis released GDPx v1.0, a benchmark of 206 long-running agentic tasks across 11 domains built with 14 partners, and found that the best model, Anthropic's Claude Opus 5.5, passed only 12.0% of tasks at $51.41 pe…
India's Ministry of Electronics and Information Technology (MeitY) released a draft handbook on 31 October 2026 extending the Data Empowerment and Protection Architecture (DEPA) to AI training, letting developers send mo…
Anthropic's Frontier Red Team, working with EPFL, MIT and StarkWare, released FlashProve, an open-source prototype that generates a cryptographic proof that a specific language model produced a specific reply. The prover…
Droneship launched a cloud runtime that runs coding agents such as Anthropic's Claude Code and OpenAI's Codex inside isolated gVisor sandboxes via a single command, `droneship run claude`, with each run receiving its own…
A developer released MnemoPay 1.15, an open-source Apache 2.0 TypeScript SDK that adds policy checks, human approvals, a double-entry ledger and tamper-evident audit chains for AI agents taking consequential actions. The…
A developer outlined an "envelope" design for unattended LLM pipelines in which the model may only choose among human-approved plans, with failure feedback carried forward as typed template messages rather than a growing…