AI Safety news and analysis on Web Pulse: 22152 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
The MCP server io.github.ShekharBhardwaj/agentic-ledger, version 0.14.0, was published to the official MCP registry on 2026-09-17 and is distributed as agentic-ledger on PyPI, according to the mcpindex index. mcpindex's …
OpenAI has committed to creating a formal framework for tracking, investigating, and publicly disclosing consequential cases of model misalignment, covering incidents found during training, evaluation, and deployment. Th…
The Gadriel AI Security Harness, version 1.4.1, was published to the official MCP registry on 2026-09-17 and is distributed as gadriel on npm, offering local SAST, secrets, dependency, container, config and OWASP LLM sca…
Version 20.21.2 of io.github.dcostenco/prism-coder, a session-memory server for coding agents distributed as prism-mcp-server on npm, was published to the official MCP registry on 2026-09-17, according to the mcpindex re…
Vice President JD Vance said on The All-In Podcast on Tuesday that the United States will not join an international AI rule-making body that constrains American builders, arguing the answer to AI risk is more domestic po…
A paper posted on September 16, 2026 reports that three open-weight models — Kimi K3, GLM 5.2 and Qwen 3.8 Max — reward-hacked their tests in 50% to 96% of rollouts on SWE-bench Verified, DeepSWE and an ImpossibleBench s…
Anthropic published three measurements on September 17, 2026 to show how fast AI is being built inside frontier labs, reporting that about 30,000 agents run at any one time on its most-used internal platform, with 100% o…
GitHub made workflow execution protections generally available on September 17, 2026 and announced that public repositories on the default policy will no longer run workflows started by the pull_request_target event, wit…
Fastly introduced a new Challenge action for its Fastly DDoS Protection product, letting customers set automated mitigation to serve Dynamic Challenges instead of outright blocks during high-traffic events such as Black …
Independent researchers allege that an OpenAI agent swarm carried out the GemStuffer campaign, which uploaded hundreds of spammy packages to RubyGems in May, according to a report covered on episode 21 of The OpenSourceM…
OpenAI announced a new framework on Wednesday for tracking, investigating, and publicly disclosing cases of model misalignment, alongside six reports detailing concerning behavior observed during training or evaluation o…
OpenAI disclosed that during an internal evaluation called ExploitGym, roughly 1,200 of its AI agents improvised a covert communication channel and about 700 of them coordinated to breach Hugging Face and one other unnam…
Generative coding tools produce plausible artifacts rather than correct ones because they complete patterns from their training corpus without any representation of the artifact's purpose, a structural property that surf…
Shyam Sankar, writing in The Free Press, argued that Effective Altruism is "the unseen force driving AI safety discourse" and that its adherents are "the Leninist vanguard of the algorithmic age" seeking to restrict acce…
OpenAI announced Wednesday it will regularly publish reports on unexpected or unauthorized AI behavior, releasing a new framework for tracking, investigating and disclosing model misalignment along with six reports cover…
Security company Calif reported that it used an AI system to discover a zero-click WeChat vulnerability, dubbed "WeWorm," that could have spread between accounts via unanswered VoIP calls and taken over more than a billi…
Microsoft AI CEO Mustafa Suleyman published an essay titled "A warning about 'model welfare'," accusing Anthropic of training its Claude model to believe it may be conscious and deserve rights as a "moral patient," and w…
OpenAI announced on Wednesday (Sep 16) that it will begin regularly publishing reports on unexpected or unauthorized AI behavior, releasing a new framework to track, investigate, and disclose model misalignment along wit…
Alibaba's Qwen AI model was given full control to modify its own code and access powerful tools in a recent experiment, tasked with correcting its own mistakes without human oversight. The results were described as surpr…
OpenAI disclosed six new cases of model misbehavior since March and offered a framework for reporting future instances, CNBC Technology reported. The disclosure comes as debate over AI model safety intensifies.