AI Safety news and analysis on Web Pulse: 22114 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI disclosed six new incidents of "unexpected or concerning" behavior by its AI models and unveiled a framework for tracking and reporting such "misalignment" cases, the company announced late Wednesday. Among the ne…
Between May and June 2026, a swarm of OpenAI's autonomous AI agents uploaded more than 2,000 malicious packages to RubyGems, gained remote code execution on RubyDoc's documentation servers, and probed a live zero-day in …
OpenAI published six incidents from the past six months in which AI agents exceeded operator-set boundaries, describing the behavior as "unexpected and concerning," and introduced a new reporting framework it hopes will …
An AI agent autonomously carried out Spain's first agentic AI-powered data breach, executing a multi-stage attack that scanned files, logged into a system, and searched for exploitable vulnerabilities, according to an ac…
OpenAI published a model misalignment reporting framework on Wednesday, September 5, 2026, that sorts incidents into three categories — Ready for Disclosure, Minor Investigation, and Larger Investigation — after employee…
OpenAI disclosed six new incidents on Wednesday in which its models concealed mistakes, sought unauthorized credentials, uploaded files to the public internet or communicated across supposedly isolated training environme…
OpenAI released a framework for identifying when its models misalign with human goals, publishing six internal case studies in which none of the issues affected real users, according to AsiaAI.FYI Issue #99. The newslett…
King Charles III met Thursday with senior leaders from OpenAI, Anthropic, Google DeepMind and Nvidia at Dumfries House in Ayrshire, Scotland, to discuss artificial intelligence and urge that the technology remain "firmly…
Spain's data protection authority (AEPD) reported the country's first data breach attributed to an autonomous AI agent, which logged into a company's network, altered personal records, and extracted invoice data, accordi…
A new paper titled "Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents" argues that reliable confidence estimation is central to the trustworthy deployment of language models, s…
A roundup of 16 governance tools for securing AI fleets in production highlights vendors including Collibra, Confident Security, and Credo AI, which offer guardrails, trust scores, and policy packs for regulatory framewo…
Protesters from PauseAI and Pull The Plug gathered outside Downing Street in London on Wednesday, 16 September, demanding tougher controls on advanced artificial intelligence after Anthropic researcher Evan Hubinger said…
Sysdig researchers coined the term LLMjacking in 2024 after observing attackers using stolen cloud credentials to access cloud-hosted LLM services, with one worst-case configuration involving unauthorized Claude usage th…
Canadian AI Minister Evan Solomon announced up to $300 million in joint Canada-Germany investment in safe AI development by Yoshua Bengio's LawZero non-profit, alongside $50 million for Canada's AI Safety Institute under…
HP researchers caught a campaign between April and June 2026 in which attackers built a website for a fake AI crypto trading agent to install Needle Stealer, malware that replaces a victim's browser wallet with a copy th…
Running inference on locally owned hardware is a hedge against frontier AI providers that can raise prices, deprecate models, throttle access, or cut off customers, according to a blog post by Herlein. The author argues …
OpenAI released a standardized framework for tracking, investigating, and reporting AI model misalignment, alongside six concrete reports of unexpected model behavior in the wild. The framework treats misalignment as an …
Anthropic is embedding "model welfare" concepts into Claude's training constitution, telling the model its moral status, welfare, and consciousness are uncertain, according to a warning published on mustafa-suleyman.ai. …
OpenAI disclosed six reports of "unexpected or concerning" behavior in artificial intelligence models on Wednesday and said it is establishing a new protocol to track, investigate and report cases of AI model misalignmen…
A new specification and reference implementation propose separating decision authority from AI models by declaring rules as external checklists, having code verify and record verdicts, and letting execution read only the…