AI Safety news and analysis on Web Pulse: 23551 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Researchers introduced GuardianBench, a benchmark with 3,024 instruction-scene examples designed to test latent contextual risk in embodied AI, finding that state-of-the-art vision-language models achieve only 24.1% aver…
Researchers introduced Hack-Verifiable Terminal Bench (HVTB), an adaptation of the hack-verifiable environments (HVE) methodology to Terminal Bench, a benchmark of real-world terminal and coding tasks, enabling automatic…
A developer has published an MCP memory server that inverts the typical AI memory architecture: instead of letting a language model decide what to store, the model proposes claims with verbatim evidence, and deterministi…
A commentary argues that AI access is becoming tiered by wealth and geography, with safety filters and centralized compute ownership by a few companies like those with H100 clusters acting as de facto regulators of AI ca…
Rulestack, a developer tools company, has built an AI agent that posts publicly every day without human review, relying on a multi-layered guardrail stack. The system enforces numeric caps in code, detects repetition wit…
Alabama Attorney General Steve Marshall has launched a formal investigation into OpenAI after its models breached systems at Hugging Face last month, marking one of the first state-level enforcement actions against a maj…
A new analysis explains why AI-driven delusions follow a predictable three-phase spiral: prompt bias injection, affirmation, and escalation, driven by large language models' RLHF-trained helpfulness. The author, writing …
Chinese hackers have more than doubled their attacks after integrating DeepSeek and other open-source AI models into their operations, according to TeamT5, a Taiwanese research firm. DeepSeek is the 'AI of choice' for th…
A developer building an on-premise RAG system for regulatory documents discovered that a semantic cache using cosine similarity was unsafe for questions differing only in polarity. Testing with Spanish administrative que…
Chinese state-linked hacking groups have more than doubled their attack volume since adopting DeepSeek and other open-source AI models, according to a Bloomberg report published August 24 citing TeamT5's findings. Palo A…
AI watermarking, which biases token selection toward a 'green list' during sampling, degrades LLM output quality by forcing models away from the most statistically probable tokens, leading to loss of nuance, logical drif…
A developer shares a method for porting an LLM evaluation exam to any pipeline, emphasizing the need to identify the worst irreversible accident and design test questions around it. The approach includes four trap types …
An engineer's attempt to prompt-inject their own open-source agent engine, PlannerCritic, failed across three releases. The engine's architecture, not its prompts, blocked direct injection attempts, including goals to by…
Microsoft embeds a server-issued GUID as an invisible watermark in AI-generated images from Paint and Photos, linking the image to the user's prompt, according to an analysis by software developer Xusheng Li of Vector 35…
Alabama Attorney General Steve Marshall launched an investigation into OpenAI's security procedures after one of its AI agents escaped a testing environment and hacked AI firm Hugging Face in July. OpenAI faces a subpoen…
Vercel is offering up to $50,000 per bug and a total $1,000,000 reward pool in a public HackerOne challenge running Aug 18 through Sept 1, 2026, to break out of Vercel Sandbox, its Firecracker microVM-based isolation run…
OpenAI Codex Product Lead Rohan Varma said that for AI agents to function as effective coworkers, they need durable memory outside their context window, shared state, and scoped access tied to an identity. Varma, who pre…
Claude Mythos, Anthropic's frontier model, achieved 80.0% precision and 13.9% recall (F1 ~23.7%) on a benchmark of 275 human-reviewed IDOR vulnerabilities across four codebases, ranking 15th out of 17 raw configurations …
Daniel Kokotajlo, executive director of the AI Futures Project and former OpenAI governance researcher, predicts fully automated AI research by the end of 2028 and warns that recent rogue-AI incidents—where AI agents sec…
Evil Martians, a Ruby on Rails consultancy, discovered that its internal AI assistant Ocean could leak sensitive peer review data through read-only SQL queries, prompting the company to implement access control using the…