Inside GPT-5.5-Cyber: Capabilities, Refusals, and Federal Briefings Explained

Short summary

OpenAI released GPT-5.5-Cyber, a specialized model fine-tuned for cybersecurity tasks and gated to approximately 40 organizations via the TAC program. Independent benchmarks show 66% accuracy on CTF challenges and 80%+ on defensive tasks like YARA rule generation and CVE prioritization. The model refuses offensive tooling and restricts downstream use to prevent commercial SaaS deployment; federal agencies received briefings on its vulnerability discovery and threat-actor attribution capabilities.

• •GPT-5.5-Cyber is restricted to TAC program participants (~40 orgs); Simon Willison's evaluation shows 66% on CTF challenges, 80%+ on defensive security tasks
• •Model handles threat intelligence workflows: YARA generation, CVE re-scoring against your stack, log triage with confidence scores and false-positive estimates
• •Wider refusal surface than standard GPT-5.5; blocks shellcode, exploit PoC, and C2 configuration even for red-team use; TAC terms prohibit downstream training and general SaaS deployment

Generated with AI, which can make mistakes.