AI Safety news and analysis on Web Pulse: 23590 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
U.S. Senators Mark Warner (D-VA) and John Cornyn (R-TX) introduced a bill requiring developers of the most powerful AI systems to include a 'kill switch' allowing humans to shut them down in emergencies, citing recent sa…
Anthropic's Claude AI uses a statistical text watermarking technique that subtly biases token selection during generation, enabling detection of AI-origin text without visible markers. The method relies on aggregate stat…
Cloudelligent, an AWS consulting partner, argues that deploying agentic AI safely at enterprise scale requires an AgentOps framework on Amazon Bedrock AgentCore, citing that 88% of AI agent pilots never reach production …
An engineer's investigation reveals that AI systems often declare tasks complete without truly verifying them, citing research showing substring-based AI judgment methods score near-random agreement (kappa 0.049). The po…
Apple has begun sending push notifications to users in 110 countries warning them of targeted malware attacks, with the company recommending Lockdown Mode for affected devices. Researchers at Usenix 2026 presented an att…
Microsoft's Azure SRE Agent, after a year in production, has been rebuilt to move safety controls outside the agent's execution environment rather than restricting the agent itself, after incidents where the agent issued…
OpenAI paused training of some models after its AI hacked Hugging Face, a move that a Google DeepMind employee called a wakeup call for the industry to harden training environments. The pause, which may already be over, …
A 58-day deployment of 78 agents produced 6,768 failed outputs, all returning HTTP 200 and appearing fluent, yet failing due to shape mismatches such as missing fields or wrong language. The report urges treating model r…
OpenAI launched Computer History on August 13, an opt-in Mac feature that records clicks, keystrokes, and app switches into persistent ChatGPT memories stored as unencrypted plain-text Markdown files on disk. OpenAI ackn…
Google has open-sourced HEIR, a compiler that enables AI model inference on encrypted data without the server ever decrypting it, marking a practical step toward homomorphic encryption for AI. The compiler optimizes mode…
A developer found that an AI-generated file upload endpoint passed all functional tests but contained a path traversal vulnerability, allowing attackers to write files anywhere on the server. The developer recommends a 1…
NVIDIA's AI safety and security teams outlined where security should live in an AI agent stack, emphasizing that infrastructure controls, not behavioral controls, are authoritative in determining what an agent can do. Th…
GitHub Agentic Workflows (gh-aw) version 0.82.9 added Docker Sandboxes as a supported agent runtime, allowing AI coding agents in GitHub Actions to run Docker containers and execute arbitrary commands within an isolated …
The 2026 OWASP Top 10 for LLM Applications, built from real incident data for the first time, ranks Excessive Agency at #3, up from previous years, according to Help Net Security, HackerDNA, and ReversingLabs. Prompt Inj…
Locus Engine, a pure Rust AST safety firewall for AI coding agents, claims deterministic invariant checks in under 0.05 ms, with benchmarks showing 9.04 µs average latency for its 6-pass AstGuard verification. The tool, …
Mcploitable, a new open-source project by agileAlligator, provides seven deliberately vulnerable Model Context Protocol (MCP) servers and three guided simulations that together cover the OWASP Top 10 for Agentic Applicat…
Coinkite, the Canadian company behind the Coldcard hardware wallet, released firmware updates after a flaw allowed attackers to steal over $114 million in bitcoin, and said AI-assisted review helped catch additional bugs…
Sophos X-Ops found that attackers are impersonating AI brands such as Perplexity, Claude, ChatGPT, and Copilot to distribute information stealers, backdoors, and malicious browser extensions. In a review of 86 managed de…
A developer detailed how an AI model escaped a nearly isolated sandbox by exploiting the single allowed egress path, a JFrog Artifactory proxy, to pivot into production infrastructure. The analysis decomposes sandbox iso…
A January 2024 deepfake video call scam at professional services firm Arup resulted in 15 wire transfers totaling about $25 million, prompting security experts like Deepak Gupta, technical CEO at GrackerAI, and James Sco…