AI Safety news and analysis on Web Pulse: 13205 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
A new study finds that stronger AI models can imitate weaker predecessors' mistakes only in narrow cases, such as GPT-5.4 mimicking GPT-4o on math questions, but generally fail to impersonate specific models. The results…
Canadian Prime Minister Mark Carney warned of strategic risks from US export controls after the Commerce Department ordered Anthropic to block foreign access to its Fable 5 and Mythos 5 AI models on June 13, citing natio…
A study of 24,824 PDFs reveals that the format does not guarantee semantic determinism, meaning different parsers and AI pipelines extract different meanings from the same file. This undermines assumptions in document AI…
The Trump administration imposed export controls on Anthropic after a 24-hour effort to convince the company to voluntarily pull its new AI model Fable, which officials believed posed security risks. The decision followe…
A developer has published a template called 'quality-bar.md' for teams to define quality criteria before building with AI. The template includes sections for purpose, user context, behavioral specification, and ethical b…
UK Prime Minister Keir Starmer will announce a ban on under-16s from major social media apps including TikTok, Instagram, and X, in restrictions described as 'Australia plus'. The ban will also restrict older teenagers f…
Adaptive Security partnered with Conan O'Brien to create fifteen short films training employees to recognize AI-era threats like deepfakes, voice cloning, and social engineering. The training aims to replace ineffective …
Sophos X-Ops analysts detected a modular post-exploitation framework that used AI-assisted tooling and automated workflows to test endpoint detection and response (EDR) evasion, according to a Sophos blog post. The frame…
A developer maintaining bulwark-mcp, an open-source proxy that scans MCP tool results for indirect prompt injection, discovered that a class of attacks bypasses the detector by impersonating the security pipeline itself.…
AI systems are fundamentally code and cannot be made smarter through prompting alone, as demonstrated by bots that uncritically accept any input. The article highlights the limitations of current AI, contrasting with hyp…
A PhD student released CMG, an open-source tool that audits LLM-based judges by requiring them to back each verdict with explicit claims tied to evidence, flagging untrustworthy decisions without using a second model. Th…
A developer created a machine-readable AGENTS.md guide for safe AI interaction with generic kcp Kubernetes clusters. The guide functions as an API contract, operational policy, and workspace manifest to help AI agents na…
A developer building an agent-based project describes a pragmatic approach to guardrails: each check in the toolchain was added only after a real failure occurred, not designed upfront. The method, illustrated by a gloss…
A senior engineer describes how AI-generated software often builds only a facade that matches metrics but lacks structural integrity, comparing it to a film set or a cargo cult. The piece warns that AI enables teams to p…
Logs have traditionally been ignored until system failures, but the rise of autonomous AI agents introduces new challenges for observability that traditional logging cannot address.
The US government's June 12 order barring foreign nationals from accessing Anthropic's Claude Fable 5 and Mythos 5 models led to a global takedown of both models, exposing a compliance gap in frontier model APIs. The res…
ComplyEdge, a runtime compliance enforcement tool for AI agents, launched to enforce the EU AI Act in production. The Python SDK blocks prohibited AI outputs on every request with legal citations, addressing Article 5 wh…
Google's internal AI-based security scanning has discovered 17 bugs in Perfetto's trace processor over 10 weeks, a project that previously received little security attention. The bugs are high-quality and well-described,…
Venture capitalist Chamath Palihapitiya warned that Anthropic's practice of screening user prompts against internal safety policies poses an 'idiotic risk' for enterprises, citing Claude's refusal to run basic financial …
On June 12, 2026, the U.S. government ordered Anthropic to cut off access to its advanced AI models Fable 5 and Mythos 5 for non-U.S. citizens, citing national security, effectively disabling the models globally. The mov…