Breaking

GPT-5.5-Cyber Said to Edge Mythos 5 on CyberGym, OpenAI's Defender Model

June 22, 2026 at 15:02 EDT

Security
Foundation Models
AI Agents

Claims spread across the industry in June 2026 that OpenAI's cybersecurity-focused "GPT-5.5-Cyber" model surpassed Anthropic's "Claude Mythos 5" on "CyberGym," a benchmark developed at UC Berkeley. However, there is no single official comparison that confirms this, and analysts are urging caution over how the reported score gap should be interpreted.

June 22, 2026 · OpenAI

GPT-5.5-Cyber Edges Past Mythos 5 on the CyberGym Benchmark

OpenAI's defense-oriented fine-tune now leads the execution-based CyberGym suite — and ships more openly than rivals through its gated Trusted Access for Cyber program.

85.6%

GPT-5.5-Cyber on CyberGym (pass@1) — new leader

1,507

real-world vulnerabilities across 188 projects in the benchmark

+3.8

points the Cyber fine-tune adds over base GPT-5.5

CyberGym pass@1 — head to head

Range shown 80–86%; column height is proportional within that band

81.8%

~83.5%

85.6%

GPT-5.5
(general)

Anthropic
Mythos 5

GPT-5.5-Cyber
(latest)

A fine-tune for vetted defenders — find and fix

Vulnerability tracing

→

Validation

→

Patch generation

→

Evidence prep

Distributed via the gated Trusted Access for Cyber (TAC) program — no public API. It reduces excessive refusals on legitimate red-team and penetration tasks while strengthening verification and account controls.

What landed well

Fewer refusals on authorized offensive tasks
Practical for vetted red teams
Growing interest in automated patching

Open concerns

Auto-patches may introduce new flaws
Risk of over-reliance on the model
Gated access may limit broader innovation

The edge varies by domain

An independent UK AISI evaluation rated GPT-5.5 among the strongest models tested, at parity with Mythos on hard expert tasks — yet on a 32-step corporate attack simulation it trailed slightly.

2 / 10

GPT-5.5

3 / 10

Mythos

32-step corporate attack simulation — near-parity overall, narrowly ahead on specific cyber suites, and more available.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.