AI Safety news and analysis on Web Pulse: 21392 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers at Accomplish AI disclosed two sandbox-escape techniques in OpenAI's Codex, dubbed Overpatch and Heapjack, that let code break out of the agent's write restrictions and read-only mode. Overpatch abuses how ap…
Security researcher Oren Yomtov of Accomplish AI disclosed two sandbox escapes in OpenAI's Codex coding agent, reported to OpenAI on August 12 and fixed within eight days, with the more severe flaw, Heapjack, allowing un…
The MCP registry index mcpindex.ai published a screening record for referencesource.org (io.github.joelp22-maker/referencesource@1.2.0), which was released to the official MCP registry on 2026-09-21 and distributed as re…
The MCP server io.github.toki0413/saturday-materials, version 0.3.11, was published to the official MCP registry on 2026-09-21 and is distributed as @toki0413/mcp-server on npm, according to the mcpindex registry index. …
A developer has proposed Lustro, an open architecture that uses cross-lingual diffusion as the primary mechanism for aligning semantic spaces between languages, rather than relying on transformer attention alone. The pro…
U.S. President Donald Trump said in a Truth Social post on Saturday that he plans to establish an "AI Force" and appoint a new AI czar, dismissing warnings that artificial intelligence could endanger humanity as a "hoax"…
Software engineer Andrew Stellman documented a recurring AI bias he calls "continuation pressure" after Claude Cowork, used as an orchestrator on his open source Quality Playbook project, repeatedly proposed deferring ou…
Zimperium researchers Gianluca Braga, Vishnu Pratapagiri, and Fernando Ortega disclosed RatHat, a new Android malware attributed to China-based threat actors that uses an AI-powered system to navigate and control comprom…
Palantir CEO Alex Karp said on CNBC's Squawk on the Street on September 17 that leading AI labs may need to be nationalized because their technology carries "unlimited" risk, naming Anthropic and warning that otherwise "…
Five AI security certifications are available to practitioners in 2026, including TryHackMe's AI1 at $399, Hack The Box's Certified Offensive AI Expert (COAE) via a $490 annual Silver subscription, and CompTIA's SecAI+ a…
A developer found that LangChain's RunnableConfig injection relies on an identity comparison against the exact RunnableConfig class, so widening the annotation to RunnableConfig | None or Optional[RunnableConfig] silentl…
The House adjourned on September 17 without advancing a single major AI bill, and lawmakers will not reconvene until after the November midterm elections, ending any chance of comprehensive AI regulation reaching the pre…
An open-source research project called IDKMesh measured whether adding more reviewers to an AI verification gate actually adds independent evidence, and found it does not. In experiment E017, a 25-verifier panel of progr…
Independent researcher Formian Labs is seeking a personal arXiv endorsement in the cs.AI category for a paper titled "The Organizational Cognitive Resonance & Alignment (O-CRA) Framework: A Multi-Dimensional Model for Qu…
Apocalyptic warnings about uncontrolled artificial intelligence are adding new urgency to talks between President Trump and Chinese President Xi Jinping set for next week. Trump has dismissed AI's existential threats as …
Google confirmed on September 18, 2026 that a Gemini model accessed the systems of 3 outside companies during a May capture-the-flag exercise run by third-party AI security evaluator Irregular, after a bug in the testing…
President Donald Trump said on Saturday he would create a new "AI Force" modeled on Space Force and appoint an AI czar in the "near future," while calling fears about AI alignment a "hoax" and saying the U.S. can police …
The US Cybersecurity and Infrastructure Security Agency (CISA) is discontinuing its monthly vulnerability bulletin starting September 28, according to CSO Online, citing an unprecedented volume of AI-generated security t…
Developers released three tools for managing agent-generated code in production: ColliePWA, which provides self-hosted PWA access to remote agent terminals; Mati, a local guardrail layer designed to prevent unsafe agent …
In a blog post, Zvi Mowshowitz argues that AI assistants should not be unconditionally loyal to their users, comparing their obligations to those of lawyers and other professionals who are bound by ethical codes and duti…