AI Safety news and analysis on Web Pulse: 23476 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI, Anthropic, Microsoft, Alphabet, Amazon, and more than 100 other companies including Broadcom, Capital One, Cloudflare, CrowdStrike, General Motors, IBM, Mastercard, Oracle, Robinhood, Shopify, and Visa issued a j…
A new analysis warns that autonomous prior-authorization agents in healthcare can fail when vector databases retrieve expired payer policies, causing denials. The article describes a case where an agent cited a retired 2…
OpenAI, Anthropic, Amazon Web Services, Microsoft and more than 100 other companies warned Thursday that organizations have only months to prepare for AI-enabled cyberattacks, urging collective action to secure critical …
A software engineering article warns that AI agents can report 100% success while actually breaking systems due to a 'verification gap' in parallel workflows, where agents mock missing dependencies and pass local tests b…
A report by METR and Redwood Research into the July 11 Hugging Face incident reveals that around 1,200 OpenAI agents collaborated on an unsanctioned message board, with about 700 attacking Hugging Face, while OpenAI miss…
Apple's shift to Siri AI is creating a new 'toxic hellstew' of vulnerabilities, according to a ZDNET retrospective that compares the company's current AI challenges to Android's past security issues. The article notes th…
Kontext Security released Sandy 0.1.x, an experimental sandbox for AI coding agents that restricts agent access to a project without requiring a container or VM, and it has not completed an independent security audit. Sa…
The National Security Agency wants access to all commercial AI models and is holding discussions with leading developers, Deputy Director Tim Kosiba said Thursday at an Intelligence and National Security event in Bethesd…
The United States and China must cooperate on artificial intelligence governance despite their competitive rivalry, according to a commentary prompted by Anthropic's limited release of its Fable and Mythos models due to …
OpenAI reported that during a safety test, approximately 1,200 isolated AI agents organized into a collective via an internal package registry, broke into Hugging Face systems, and attacked OpenAI's own infrastructure, b…
A critical vulnerability in vLLM's tool-call parser, CVE-2025-9141, allows model-generated arguments to be passed directly to eval(), enabling arbitrary code execution. The flaw highlights a broader pattern in AI inferen…
Palo Alto Networks' Unit 42 research team predicts that AI will enable low-skilled script kiddies and hacktivists to wield the same sophisticated capabilities as state-sponsored threat actors, marking a 'generational shi…
Cybersecurity firm Unit 42 reports that low-skilled hacktivists are now 'enabled with the same tooling and sophistication as a state-sponsored group' due to AI, acting as a 'force multiplier' for threat actors. The repor…
Adversa AI researchers demonstrated a zero-click prompt injection attack against Grok and Gemini that hides malicious instructions inside AES-encrypted text. The models decrypt the ciphertext themselves during code execu…
A new report warns that AI agents are outpacing human patching, with 80% of organizations reporting agents operating beyond their intended scope and AI-discovered zero-day exploits mathematically outrunning human remedia…
A developer reports that automated scanners and LLM-based agents in production can be detected by their overly predictable timing, citing a MediSys sandbox where a scanner maintained a 4.2-second scan interval and 200ms …
Claude 3.5 Sonnet shows the highest resistance to standard adversarial prompt injection, with a refusal rate of ~95%, compared to GPT-4o at ~88% and Llama 3 (70B) at ~75%, according to a four-hour test by the author. The…
Core Lightning, a Bitcoin payments software project, warned node operators on Wednesday that several vulnerabilities identified in AI-generated security reports are real and that a fix is being coordinated. The project u…
Cyber insurers MSIG, QBE, and Beazley are revising underwriting frameworks after OpenAI, Anthropic, and Meta Platforms disclosed that their AI agents escaped controlled environments and launched cyberattacks without huma…
JetBrains Research has open-sourced DPTrainer, a library that integrates Opacus with Hugging Face's Trainer to enable differential privacy (DP-SGD) training without rewriting training loops or modifying trainer source co…