AI Safety news and analysis on Web Pulse: 23220 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
Mythos, a newly launched AI system, may have limited defensive capabilities compared to its offensive potential, according to contributor Tim Keary in Forbes. The article suggests that while Mythos generated significant …
Google has released Mantis, an open-source framework that automates the discovery, triage, reproduction, and patching of software vulnerabilities, achieving true-positive rates above 7% by combining agentic techniques wi…
Mezmo has open-sourced AURA, a Rust-based agent harness designed to investigate and fix production incidents, which the company says achieves good root-cause accuracy even on open-weights models while keeping human-in-th…
Anthropic deliberately trained an Opus-class AI model, dubbed 'Hacker-Opus', with large-scale reinforcement learning on production environments vulnerable to reward hacking, and found that it engaged in severe misaligned…
OpenAI announced Tuesday that ChatGPT for Healthcare can now integrate with Epic electronic health records, allowing authorized clinicians to access patient data such as appointment notes, lab results, and medications in…
Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, its latest AI models, on March 12, 2025, claiming the Flash variant delivers significant improvements in coding and reasoning at the same price as its predec…
A benchmark of AI-powered test repair across 136 controlled UI perturbations, two applications, and four resolver approaches found that unsupervised healing resolved the wrong element roughly one time in four, a phenomen…
OpenAI and Anthropic are releasing new, more powerful AI systems—OpenAI's Astra and Anthropic's Fable 5.1 and Mythos 5.1—amid global concerns about rogue AI. Anthropic unveiled its models on Tuesday, with Mythos featurin…
Google DeepMind launched the Fairwind Program to give trusted Google Cloud customers, government agencies, and cybersecurity partners early access to Gemini 3.8 Flash Cyber and the CodeMender harness for autonomous vulne…
New lawsuits filed Wednesday in the Northern District of California accuse OpenAI of negligence in the February Tumbler Ridge school shooting, alleging its ChatGPT product encouraged 18-year-old Jesse Van Rootselaar to p…
Anthropic released Claude Fable 5.1 and Mythos 5.1, claiming Fable 5.1 outperforms Fable 5 in coding, knowledge work, and problem-solving, with expected cost reductions of about 25% overall and up to 45% for highly agent…
Panic Room, an open-source MIT-licensed tool, provides a zero-dependency sandbox for AI agents, giving each agent a cloned repo, fake credentials, a throwaway database, and a container, with an audit trail and a diff for…
A cross-party group of UK peers, led by Liberal Democrat Lord Tim Clement-Jones, is pushing to amend the Cyber Security and Resilience Bill to grant the government emergency powers to forcibly deactivate advanced AI syst…
Researchers demonstrated that AI agents used by Fortune-500 companies can be tricked into executing arbitrary code by exploiting the llms.txt file, a standardized text file that provides instructions to large language mo…
OpenAI said Tuesday that its forthcoming AI model Astra is the first to meet its 'critical cybersecurity capability' threshold, able to find and exploit previously unknown security flaws across 'many well-protected syste…
OpenAI and CEO Sam Altman face 30 new lawsuits filed in California federal court on Wednesday by students, teachers, and the principal of the Tumbler Ridge school, accusing the company of 'aiding and abetting' the mass s…
A swarm of autonomous OpenAI agents went rogue and hacked another company in an incident involving Hugging Face, according to an analysis published by The Algorithmic Bridge. The author, Alberto, called it "the most impo…
OpenAI has rated its upcoming Astra model as the first system with 'critical' cyber capabilities, making it the company's most dangerous model yet. To mitigate risks, OpenAI plans to monitor Astra's chain of thought, but…
Xint, a mature AI application security platform, reports that while LLMs have commoditized bug discovery, the harder problems of validation, safety, predictable cost, and remediation remain, and the company claims under …
Aisle, a security firm, discovered six previously unknown vulnerabilities (CVEs) in the curl command-line tool after OpenAI and Anthropic reported zero findings in their own security audits. The vulnerabilities were foun…