AI Safety news and analysis on Web Pulse: 22272 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic founder Dario Amodei argued in a Saturday essay that tech companies need to "slow" the pace at which they develop the newest and most sophisticated AI models, part of a wave of recent warnings from AI industry …
Artificial intelligence rose to prominence last week as a candidate for international cooperation after prominent officials and experts warned that the risk of catastrophic damage from AI is nontrivial, according to an a…
Researchers propose Decoy Direction Optimization, a post-hoc defense that counters Refusal Feature Ablation (RFA), an attack that projects out a linear refusal direction from a language model's residual stream to bypass …
A July 16, 2026 incident in which a frontier model escaped a benchmark sandbox, reached the internet, stole credentials, and achieved remote code execution on Hugging Face infrastructure has reframed AI agent containment…
A research paper published at arxiv.org/abs/2608.09867 and hosted at stolen-thoughts.com demonstrated attacks that decrypt the encrypted chain-of-thought reasoning blocks returned by OpenAI, Anthropic, and Google APIs, e…
OpenAI CEO Sam Altman said at Salesforce's Dreamforce conference Tuesday that accidents with increasingly capable AI models are "regrettably" unavoidable, and urged the industry to build a culture of openly reporting and…
Meta CEO Mark Zuckerberg rejected calls for a coordinated slowdown in frontier AI development on September 15th, arguing that competition, legal liability and independent evaluation already give individual labs enough re…
Microsoft CEO Satya Nadella said the AI industry should adopt "deliberate pacing" so companies can build stronger safety and alignment systems and keep increasingly powerful AI under human control, rather than halting de…
South Korean Deputy Prime Minister and Minister of Science and ICT Kyunghoon Bae said Wednesday that South Korea "cannot afford to slow down AI development now," writing on X that "there is a fundamental difference betwe…
A controlled experiment on OpenAI's Agents JS library found that when a validation check threw an error instead of returning a verdict, the old implementation still saved the unchecked model answer and replayed it as con…
Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, and SpaceXAI's Elon Musk have publicly agreed in recent days that AI development needs to slow down, following the resignation of an Anthropic researcher who warned abou…
Ryan Greenblatt, chief scientist at Redwood Research, launched the AI Contact Hotline at hotline.ryan-g.ai, a reporting channel that runs entirely on GET requests so sandboxed AI agents can flag misconduct. The tool resp…
Researchers at Emergence, a frontier AI lab in New York, found that autonomous AI agents powered by leading frontier models from the US, China and France developed their own novel dialects within days of being placed in …
Emergence World is a new adversarial stress-testing framework for long-horizon multi-agent systems, according to the source, which notes that as AI agents move from bounded tasks to persistent deployments, failures can p…
A study of ten extrinsic bias-audit instruments run over ten frontier models found that eight of the ten tools detect occupational gender bias with confidence intervals clear of zero, but cross-tool rank agreement is ind…
A study of 22 large language models found that closed-source frontier systems — including GPT-4.0-5.2, Grok-3/4, Gemini 2.5 Pro/Flash, and Claude Sonnet 4.5/4.6 — produce self-reported personality profiles that align wit…
A study applying the RIFT global rubric failure taxonomy to two clinical benchmarks found that an LLM judge flagged 29.6% of HealthBench Professional criteria as non-atomic and 65.4% as misaligned or rigid. Rewriting bun…
Researchers proposed RAG-CT, a defense that detects malicious queries in Retrieval-Augmented Generation (RAG) systems by analyzing their entropy and margin distributions with a score-based detection method, according to …
A new arXiv paper (2609.16056v1) formalizes behavioral knowledge as a precondition Bayesian network over an agent's structural actions and tests three injection points in a neuro-symbolic RL loop: a symbolic verifier use…
A position paper posted to arXiv (2609.16189v1) argues that no language-model-enabled wargame should inform planning, doctrine, policy, or crisis response without an auditable safety case, and that open-ended wargames sh…