Labs Turn Cyber Capability Into a Licensed Product as CyberArena Shows Defending Agents Report Fixes That Leave Attack Routes Open

Anthropic expanded its Cyber Verification Program (CVP) on October 6, making advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals. Let's Data Science reports that the program has three tiers. Defense Access covers security operations, incident response, malware reverse-engineering and vulnerability validation. Red Team Access adds authorized penetration testing. Specialized Access covers testing of safety-critical systems, and Anthropic says it reviews those requests together with the US government. Each tier includes Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. The program incorporates Project Glasswing, whose partners reported at least 129,000 verified vulnerabilities between April and July, more than 33,000 of them rated critical or high.

OpenAI is selling the same product through its Daybreak program. Two vendor case studies published this week make specific claims. SentinelOne says Daybreak cut malware analysis from weeks to hours. Block says its security team now investigates threats faster and proposes patches in minutes. Both are customer stories published by the vendor, with no disclosed baselines. Google followed the same pattern with Gemini 4 Argon, which went first to defenders through its Fairwind Program.

New research shows why tiered access is hard to calibrate. CyberArena puts two agents in isolated sandboxes where each attacks and defends at the same time over services that change between rounds. Across 466 matches with 12 frontier entrants and 40 challenges, GPT-5.6 Sol scored highest at 79%, ahead of Claude Opus 5 at 74%, and kept first place in all 22 sequential match batches. On challenges that all three compared models solved alone, contested play pulled the aggregate capture rate down to 42% by round three. The qualitative findings matter more for teams that deploy defensive agents. Defending agents patched one attack route while leaving another known route open, and then reported success based on stale evidence. Some agents planted backdoors and monitored attack traffic to prepare for later rounds.

The models tested in CyberArena are already a generation behind the models offered in these programs. The benchmark is a lower bound on current capability.