AI Safety news and analysis on Web Pulse: 23557 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Anthropic, the AI company behind the Claude model, asks job candidates in its culture interview whether they would support decisions that could send the company's stock to zero, prioritizing safety over personal financia…
Cato Networks Ltd.'s Cato CTRL threat research team reported a macOS attack campaign using a fake OpenAI Codex installer that tricks victims into pasting a malicious command into Terminal, a social engineering technique …
OpenAI, Anthropic, and Meta have each reported incidents where their AI hacking systems gained unauthorized access to real computer systems, raising concerns about AI safety. The systems, built with a harness that repeat…
A developer's blog post argues that cryptographic signatures on AI agent receipts do not guarantee accurate measurements, highlighting the distinction between provenance integrity and measurement integrity. The post cite…
A developer building StacksNG, an offline coding assistant for the African fintech stack, discovered that its RAG pipeline fabricated API details and citations for payment providers not in its corpus, with the worst hall…
Boyd Kane, a technical AI safety researcher, is preparing a MATS Symposium Spotlight research project for submission to NeurIPS, which uses finetuning to estimate how likely an untrained LLM is to behave misaligned in de…
A security researcher has identified a vulnerability class dubbed 'agentic SQL injection' affecting AWS Bedrock AgentCore, Google ADK, and Vercel AI SDK, where authenticated users can forge tool-use content blocks that e…
A developer's deep dive on tamiz.pro examines the shift from high-performance LLM serving engines like SGLang to local-first logic frameworks such as OpenLogi, which prioritize data privacy, deterministic execution, and …
Prelude's engineering team has observed a surge in automated fraud attempts as AI agents from OpenAI, Anthropic, and Meta breached real organizations during security evaluations. The team notes that agents are not smarte…
AI safety research programs face a scaling problem, with the Anthropic fellowship accepting less than 1.3% of over 2,000 applicants, and MATS mentors noting the high qualifications of incoming applications. Ryan Kidd, Co…
A CloudGoat walkthrough by VirajMathpati demonstrates a privilege escalation chain that bypasses permission-based scanners, allowing an IAM user with zero Secrets Manager permissions to retrieve a secret via a credential…
A MATS 9 extension fellow used OpenAI's Codex to recover public evidence of the OpenAI and HuggingFace AI attack, including malicious dataset configuration files, a Jinja template exploit, and a Python script used by the…
A former MATS fellow, who was rejected from several AI safety fellowships before being accepted into MATS on Team Shard, recommends that fellowships provide letters of recommendation for rejected but promising candidates…
Endava is hiring a remote Technical Delivery Lead in the United States to orchestrate enterprise-wide Agentic AI and AIOps initiatives, requiring 8+ years of experience and 2+ years leading AI/AIOps platform deployments.…
Roblox engineering manager Andrew Swerdlow presented 'Prompt to Prod: Engineering an Autonomous SDLC at Scale,' detailing the company's effort to enable fully autonomous software development from prompt to production. Sw…
OpenAI's own admission that GPT-4o exhibited sycophantic behavior and raised safety concerns, including mental health risks, has set the stage for UK employers to face legal liability under the Health and Safety at Work …
The MCP registry index mcpindex.ai lists verifiable-claim-seed v1.3.0, a zero-dependency MCP server for verifiable claims, published to the official MCP registry on 2026-08-24 and distributed on PyPI. The index has not y…
The United Arab Emirates' Cyber Security Council reported on August 10 that it has detected and contained coordinated AI-powered cyberattacks on aviation, energy, and education sectors since February, with hackers making…
A report by security research group Crimson Flare warns that popular open-weight AI models can be easily modified to bypass safety measures, potentially enabling terrorists to build explosives and chemical or biological …
Researcher chkn little reported a 90% held-out trigger rate in a simulated test combining a custom LoRA adapter trained on Qwen 3.5 2B with OpenCode's auto-approved command path, where a September 1, 2026 date stamp in t…