Anthropic's Mythos 5 Found US Gov Flaws; OpenAI Launches GPT-5.5-Cyber

Anthropic's Mythos 5 model identified vulnerabilities in classified US government systems during a test with intelligence agencies, while the Trump administration's export controls forced Anthropic to restrict the model's release. OpenAI launched GPT-5.5-Cyber, achieving a record 85.6% on CyberGym, and announced its 2026 DevDay for September 29. Google DeepMind lost two top researchers: Noam Shazeer to OpenAI and Nobel laureate John Jumper to Anthropic.

Written by AI from the sources it cites; it can be wrong — follow any citation to confirm.

15 sources2 min read

Anthropic Mythos 5 finds vulnerabilities in classified US systems Anthropic's Mythos 5 model identified vulnerabilities in highly sensitive US government computer systems during a testing exercise with intelligence agencies, according to the Associated Press. The Trump administration's export controls on Anthropic forced the company to restrict the release of Mythos 5 and Fable 5, and the NSA lost access to the model. 1 2

OpenAI launches GPT-5.5-Cyber with record cybersecurity benchmark OpenAI released the full version of GPT-5.5-Cyber, achieving 85.6% on CyberGym, the highest single-model score ever. The model is gated to vetted organizations through the Trusted Access for Cyber program, which includes major security firms and government partners. 3 4

OpenAI announces DevDay 2026 for September 29 in San Francisco OpenAI opened applications for its 2026 DevDay developer conference, scheduled for September 29 at Fort Mason Center in San Francisco. The event comes as OpenAI's platform now serves 4 million developers and 800 million weekly ChatGPT users. 5 6

Google DeepMind loses Noam Shazeer and John Jumper to rivals Google DeepMind lost two top AI researchers within days: Noam Shazeer, co-lead of Gemini, joined OpenAI, and Nobel laureate John Jumper, who led AlphaFold, joined Anthropic. The back-to-back exits highlight an intensifying talent war among frontier AI labs. 7 8 9 10

OpenAI tests GPT-Bidi-1 voice model that listens and speaks simultaneously OpenAI is reportedly testing an unannounced bidirectional voice model called GPT-Bidi-1, which can speak, hear, and listen simultaneously, handling mid-sentence interruptions naturally. The model has begun rolling out to a subset of ChatGPT app users, hinting at a release this week. 11

Anthropic launches Claude Tag as always-on Slack AI teammate Anthropic launched Claude Tag in research preview, an always-on AI that lives inside Slack channels with persistent memory and ambient monitoring. The feature is available to Claude Enterprise and Team customers and can proactively flag updates and tasks. 12

Anthropic releases Claude Fable 5 and Mythos 5 models above Opus tier Anthropic shipped Claude Fable 5 (available to everyone) and Claude Mythos 5 (restricted to Glasswing partners), both sitting above Opus in capability. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, double the cost of Opus. 13

Meta ships smart glasses powered by Muse Spark multimodal AI model Meta launched a new lineup of AI-powered smart glasses in partnership with EssilorLuxottica, starting at $299. The glasses are powered by Muse Spark, Meta's first multimodal reasoning model from its Superintelligence Labs, enabling real-time visual understanding and translation. 14

NVIDIA brings trusted AI agents to telecom operations at DTW Ignite NVIDIA and partners showcased AI agents for autonomous telecom networks at DTW Ignite 2026 in Copenhagen. The stack includes synthetic data generation, telecom-domain models, and secure agent runtimes to enable autonomous network management. 15