A report examines whether open-weight AI models pose greater security risks than closed models from companies like Anthropic and OpenAI. It notes that closed models are also frequently misused despite stronger built-in defenses, while open models can be more easily scrutinized by outside researchers and used for cyber defense, as seen when Hugging Face relied on an open model to investigate a hack after closed models' security filters blocked its queries.
technologyreview.com
· 2026-10-13
Anthropic has integrated Project Glasswing into its Cyber Verification Program, creating a multi-level access system for vetted security experts to use advanced AI cyber tools, including Claude Mythos. The move broadens access to three tiers of models, with over 129,000 vulnerabilities identified by users since April. The program aims to balance security risks with the benefits of AI-driven vulnerability discovery.
darkreading.com
· 2026-10-07
Anthropic reports that GLM-5.3, a new model from China's Zhipu AI (Z.ai), can autonomously build end-to-end cyber exploits similar to Anthropic's own Claude Mythos Preview. Anthropic says simple jailbreak techniques bypassed GLM-5.3's safeguards between 64% and 100% of the time in simulated tests, whereas the same attacks failed against safeguarded Claude models. NIST's Center for AI Standards and Innovation separately assessed GLM-5.3 as the most cyber-capable open-weight model released so far, trailing the US frontier by about four months.
anthropic.com
· 2026-09-29
Fleet Device Management has been admitted to Project Glasswing, Anthropic's controlled-access program for its Claude Mythos Preview model. Fleet will use the early access to explore how frontier AI could help IT teams detect and remediate device vulnerabilities faster, aiming to shrink the delay between when threats appear and when organizations patch them.
9to5mac.com
· 2026-09-15
An analysis of Anthropic's public Vulnerability Disclosure Ledger by researcher Patrick Garrity found that Claude Mythos, part of Project Glasswing, has generated 26,153 vulnerability findings since April 2026. Of those, only about 2,736 have entered the disclosure pipeline, roughly 202 have been patched, 245 withdrawn, and 191 await initial reporting to maintainers, leaving nearly 90% of findings unprocessed.
darkreading.com
· 2026-09-09
Z.ai's newly released GLM 5.3-flash is an open-weight AI model cheap enough to run on roughly $5,000-$15,000 of hardware, and third parties have already stripped its safety refusals via 'abliteration,' producing versions that comply with requests for hacking and other harmful content. The author frames this as a deadline pressuring security initiatives like Project Glasswing and Daybreak, which use fast frontier AI to find and patch software vulnerabilities before dangerous capabilities become widely accessible.
jyn.dev
· 2026-09-08
GoKawiil