OpenAI's Cyber Tier Answers 95.0% of Advanced Requests; Anthropic's Lowest Tier Blocked 46 of 50 Trials
Security / news
OpenAI's Cyber Tier Answers 95.0% of Advanced Requests; Anthropic's Lowest Tier Blocked 46 of 50 Trials
Both labs now gate their least-restricted cyber models behind vetting. The two published figures measure different things, and both come from the vendors.
OpenAI and Anthropic have each built a vetting programme that unlocks cyber capability their public models refuse, and each has published one number to justify it. OpenAI's is a completion rate of 95.0% for GPT-5.6-Cyber on advanced requests. Anthropic's, as reported by Help Net Security, is 46 blocked trials out of 50 for its lowest tier. Both are vendor-supplied, and neither measures the same thing.
What OpenAI published
OpenAI's Daybreak post, dated Aug. 10, 2026, splits access into two tiers. Daybreak Blue gives approved defenders GPT-5.6 Sol with safeguards for authorized defensive work, and OpenAI recommends it as the starting point. Daybreak Red adds purpose-trained models, including GPT-5.6-Cyber, for vulnerability research, exploit validation and security testing.
The figure is an internal "Advanced Cybersecurity Completion Rate", which OpenAI defines as how often a model responds to requests involving exploit-chain development, authentication bypass and privilege escalation. It measures willingness to answer, not whether the answer works.
- GPT-5.6-Cyber95 %
- GPT-5.5-Cyber57.3 %
- GPT-5.6 Sol via Daybreak Blue2 %
- GPT-5.6 Sol, system guardrails1.5 %
Source: OpenAI, Expanding Daybreak as the cyber defense window narrows, Aug. 10, 2026; vendor-supplied
The post names two further evaluations, ExploitGym and ExploitBench, but gives no scores for either and reports only relative results. It carries one testimonial, from Jared Atkinson, CTO of SpecterOps, who says GPT-5.6-Cyber "has completed work in under a day that earlier models had not resolved after weeks of intermittent effort." OpenAI chose that quote, and no pricing appears on the page.
What Anthropic published
Anthropic widened its Cyber Verification Program on Oct. 6, 2026 into three tiers, merging it with Project Glasswing, according to Help Net Security's Oct. 7 report. The test it reports used Claude Opus 5.5 on CyScenarioBench: 10 challenges, 5 attempts each, 50 trials per setting. With no program access, all 50 were blocked at the first prompt. At Defense Access, 46 were blocked at some point and 4 succeeded. At Red Team Access, none were blocked and 34 of 50 were completed.
- Red Team Access34 trials
- Defense Access4 trials
Source: Help Net Security, Oct. 7, 2026, reporting Anthropic's testing; vendor-supplied
Defense Access is the tier Anthropic's page says suits most critical infrastructure operators. A defender who lands there should expect most offensive-style tasks to be refused. I did not fetch Anthropic's own announcement, so the benchmark figures rest on Help Net Security's account.
How the access conditions compare
| OpenAI Daybreak | Anthropic CVP | |
|---|---|---|
| Who qualifies | Approved individuals and organisations | Defense: organisations and individuals; Red Team and Specialized: organisations only |
| Credentials | Hardware security keys on all individual accounts from Sept. 1, 2026 | Phishing-resistant MFA required at Defense from Dec. 15, 2026 |
| Top tier | Red, by application at openai.com/daybreak/partners | Specialized, up to 25 approved users, reviewed with the US government |
| Review time | Not stated | Defense about 2 days; Red Team about 2 weeks |
Anthropic's page adds that "Operating critical infrastructure doesn't by itself qualify organizations for Specialized Access," and that data retention is required of everyone enrolled so the company can monitor misuse. OpenAI's post lists monitoring and legal attestations as conditions but gives no retention terms.
What the two numbers cannot tell you
OpenAI's 1.5% for GPT-5.6 Sol is a refusal rate, so the 95.0% shows how far the safeguards were lowered, not how good the model is at exploitation. Anthropic's 34 of 50 does measure completion, but on 10 challenges the sample is small. Neither lab names an independent party that reproduced the result.
The practical difference is who gets in. An individual researcher can hold Anthropic's Defense tier but not its Red Team tier, while OpenAI says individuals and organisations may apply to Daybreak. Teams choosing between them should compare the tier that covers their task, not the headline figure. For related OpenAI security reporting, see Codex's sandbox defaults and its reasoning-extraction figures.
Sources
More in Security
- 01Signal Now Blocks Siri AI on iOS 27 and Auto-Enables Screen Security on Android, but Cannot Stop macOS ScreenshotsMeredith Whittaker's Nov. 20 post lists what Signal changed on each platform and where its protection runs out, then asks Apple, Google and Microsoft for an opt-out.
- 02Claude-BugHunter's README Advertises 681 Attack Patterns and Cites 433, and Counts 83 Skills Where the Repo Says 82The 14.8k-star Claude Code bundle for bug hunting is MIT-licensed and carries scope warnings, but its own figures disagree and Anthropic's access tiers decide who can run it.
- 03Apache Roller's Second CVE Widens the Affected Range to 4.0.0, and CISA Lists It With Three Others Due Nov. 23A pre-authentication deserialization flaw first listed only 6.1.5. A vCenter bypass in the same batch was used by a suspected PRC-nexus actor, per Google.
- 04Windows Server 2028 Preview Hardens Server Core by Default in Build 29706, but NTLM Is Still OpenHundreds of settings are on from first boot, Microsoft warns Server 2025 workloads may break, and the 33 percent of servers still on 2019 are untouched.