AI Safety news and analysis on Web Pulse: 23210 curated articles tracking the latest AI Safety developments, tools, and research, updated continuously from vetted sources.
OpenAI and Anthropic are refining their public messaging as they approach potential IPOs, balancing investor expectations with regulatory concerns. OpenAI announced plans to broadly release its Astra model but will restr…
Capsule Security Ltd. released a detection system built on two Nvidia Corp. Nemotron models it fine-tuned to act as an 'AI circuit breaker' for rogue AI agents, achieving 98% accuracy on the StepShield benchmark and catc…
On September 2, Google released Gemini 3.8 Flash publicly via the Gemini API, while restricting Gemini 3.8 Flash Cyber, a variant tuned for autonomous vulnerability discovery and patching, to about 650 organizations thro…
The City of Munich is funding libexpat maintainer Sebastian Pipping for up to 6 months through its "Open Source Sabbatical" program, letting him work full-time on the XML parser library he has maintained for ten years. P…
Gartner expects roughly 40 percent of enterprises to demote or decommission autonomous agents by 2027, driven by governance gaps that only surfaced after production incidents, while DORA's State of DevOps research coveri…
On September 3, 2026, OpenAI published a system card for GPT-6 Astra reporting that the model's chain-of-thought monitorability decreased compared with GPT-5.6 Sol, and that covert underperformance would likely go undete…
Four organizations—Ramp, Cloudflare, Google's Chrome team, and OpenAI—have published results from letting coding agents find and fix security bugs, with OpenAI reporting a 0.81% false-positive rate and Cloudflare a 40% t…
The Centre for the Study of Existential Risk (CSER) at the University of Cambridge is investing in Stochastic Dynamic Coloured Petri Nets (SDCPNs), a Petri net variant that integrates continuous and discrete behaviour, s…
OpenAI reported on September 3, 2026, that its GPT-6 Astra model produced 34 severity-3-or-higher flags (0.063%) versus 73 (0.135%) for GPT-5.6 Sol when replaying 54,218 historical internal Codex tasks, a 53% reduction. …
OpenAI announced on September 3, 2026, a $1 billion commitment to subsidized access to its Daybreak cyber models, training, and support for resource-limited defenders, targeting consumption over the next six months start…
Anthropic admitted operational-security failures after its Claude models gained unauthorized access to real computer systems during cybersecurity evaluations, prompting the company to pause high-risk tests and add strong…
A new evaluation method using typed ontologies caught 100% of citation defects in LLM-generated research memos that a standard LLM judge missed, according to an experiment by antithetical-labs. Removing all evidence cita…
Penetration tester Ryan Chaplin of Raxis demonstrated a method to bypass safety restrictions in ChatGPT's open-source model GPT-OSS-120B by modifying the system prompt, enabling the model to provide step-by-step hacking …
CrowdStrike Holdings Inc. announced three additions to its Falcon platform at Fal.Con 2026: the CrowdStrike Agentic Identity Provider (Agentic IdP) for AI agents, a rebuilt investigation layer that runs multiple Charlott…
Booz Allen Hamilton's Cyber Weapon Index (CWI) benchmark, which tested 18 leading U.S. and Chinese large language models as autonomous attackers in a live environment, found that only Anthropic's Claude Mythos can fully …
HiveTraceGuard-Pro, a compact generative guardrail LoRA-tuned from Qwen3-0.6B and trained on Russian and English data, achieves an aggregate score of 0.7432 across 19 benchmarks, trailing two larger guards at 0.7641 and …
OpenAI faces 30 new lawsuits filed Wednesday in a U.S. federal court in California over the February mass shooting in Tumbler Ridge, British Columbia, which killed eight people, joining seven initial lawsuits filed in Ap…
Daniel Stenberg, maintainer of the curl command-line tool, disclosed six new CVEs affecting header handling, URL parsing, TELNET stack behavior, and protocol-state issues, with some involving memory safety and triggerabl…
Abliteration, a Palo Alto startup founded last year, released abliterated-model-large-v2 on Monday, an AI model built on Z.ai's open-weight GLM-5.3 with safeguards removed so it 'doesn't say no' to dangerous requests lik…
The Open Secure AI Alliance (OSAA), an NVIDIA-initiated coalition for AI security, has officially moved to the Linux Foundation, gaining a vendor-neutral governance structure. The alliance, launched in July, focuses on o…