Sam Altman Eats His Words on Cyber Access
Two months ago Sam Altman called Anthropic's restricted Mythos rollout 'fear-based marketing.' On April 30 OpenAI announced GPT-5.5-Cyber will ship to a vetted, government-coordinated list of 'critical cyber defenders' before anyone else gets it. Same playbook, different logo.
Cyber is the cybersecurity-tuned variant of GPT-5.5. It does autonomous penetration testing, vulnerability discovery, malware reverse engineering. Anthropic Mythos was already doing zero-day discovery across browsers and operating systems under Project Glasswing. OpenAI's pitch as recently as April 9 was 'we ship to everyone, that's how science moves' — Altman literally tweeted that. April 30 he's on stage saying access will expand 'only after consulting the U.S. government.'
The technical content matters less than what just happened to the agent security category. Both frontier labs now agree that the most capable cyber agents need gated access. That settles a debate that's been running since Project Glasswing in early April. Vendor selection for agentic SOC products (CrowdStrike Falcon, Microsoft Sentinel Agentic, SentinelOne Purple AI, Arctic Wolf Aurora) just got constrained — none of those vendors get the unrestricted top-tier model from either lab. The model itself is now a procurement gate.
Second-order: this is the first time OpenAI publicly admits an agent capability is dangerous enough to gate. Six months of the AGI-soon narrative was built on 'safety as marketing' framing. The Cyber climbdown is the first crack — not because OpenAI changed its mind, but because the U.S. government got a vote.
Full TechCrunch story at https://techcrunch.com/2026/04/30/after-dissing-anthropic-for-limiting-mythos-openai-restricts-access-to-cyber-too/
← Back to all articles
Cyber is the cybersecurity-tuned variant of GPT-5.5. It does autonomous penetration testing, vulnerability discovery, malware reverse engineering. Anthropic Mythos was already doing zero-day discovery across browsers and operating systems under Project Glasswing. OpenAI's pitch as recently as April 9 was 'we ship to everyone, that's how science moves' — Altman literally tweeted that. April 30 he's on stage saying access will expand 'only after consulting the U.S. government.'
The technical content matters less than what just happened to the agent security category. Both frontier labs now agree that the most capable cyber agents need gated access. That settles a debate that's been running since Project Glasswing in early April. Vendor selection for agentic SOC products (CrowdStrike Falcon, Microsoft Sentinel Agentic, SentinelOne Purple AI, Arctic Wolf Aurora) just got constrained — none of those vendors get the unrestricted top-tier model from either lab. The model itself is now a procurement gate.
Second-order: this is the first time OpenAI publicly admits an agent capability is dangerous enough to gate. Six months of the AGI-soon narrative was built on 'safety as marketing' framing. The Cyber climbdown is the first crack — not because OpenAI changed its mind, but because the U.S. government got a vote.
Full TechCrunch story at https://techcrunch.com/2026/04/30/after-dissing-anthropic-for-limiting-mythos-openai-restricts-access-to-cyber-too/
Comments