OpenAI Rolls Out GPT-5.5-Cyber to Vetted Defenders, a Month After Mocking Anthropic's Mythos as 'Fear-Based Marketing'
OpenAI launched GPT-5.5-Cyber on May 7, 2026 to vetted security teams via its Trusted Access for Cyber program, after CEO Sam Altman publicly criticized Anthropic's restricted Mythos rollout.
Editor's Note ·
- Clarification:
- The automated review flagged the phrase 'the most clearly malicious workflows' in the Analysis section as 'loaded language'. In editor judgment this is a false positive — 'clearly malicious' is used as a measured qualifier denoting an unambiguous category (credential theft, attacks on third parties, which the article previously specifies) rather than editorializing. The article publishes as written; the flag is recorded here for audit transparency.
- Clarification:
- Axios source was fetched via Archive.org fallback rather than from axios.com directly. Content was verified to match the live article. This is editorial-process metadata, not a factual issue.
Overview
OpenAI on May 7, 2026 began rolling out GPT-5.5-Cyber, a variation of its latest frontier model trained to be more permissive on security-related tasks, to a small group of vetted cybersecurity teams, The Decoder and CNBC reported. The release lands roughly a month after Anthropic gave a similarly restricted preview of Claude Mythos to a handpicked group of partners — and only days after OpenAI CEO Sam Altman dismissed that approach as “fear-based marketing,” TechCrunch reported.
In a statement carried by Help Net Security, Altman said: “We’d like to help companies secure themselves, and we think it’s important to start work on this quickly.”
What We Know
A three-tier access model
GPT-5.5-Cyber sits at the top of a new three-tier permissions structure that builds on the Trusted Access for Cyber framework OpenAI introduced in February 2026, as previously reported. According to The Decoder, there is now a public model with standard restrictions, a middle tier with relaxed filters for defensive work, and GPT-5.5-Cyber with the fewest restrictions for authorized penetration testing. Help Net Security describes the new top tier as “the most permissive version in its cybersecurity lineup,” while noting that GPT-5.5 with Trusted Access remains the recommended entry point for most workflows.
According to TechCrunch, the Trusted Access for Cyber, or TAC, program has scaled “to thousands of verified defenders and hundreds of teams responsible for protecting critical software.”
Launch partners
The Decoder reports that launch partners include Cisco, CrowdStrike, Palo Alto Networks, Cloudflare, Intel, Snyk, and SentinelOne. Select developers on major open-source projects are also receiving discounted access through a program called Codex Security, the outlet reports.
What the model is allowed to do — and what it is not
Approved tasks for GPT-5.5-Cyber include authorized red teaming, penetration testing, and controlled vulnerability validation, according to Help Net Security. The Decoder lists adjacent capabilities such as malware analysis, security-patch review, reproducing known vulnerabilities for patching, and penetration testing.
Guardrails remain. The Decoder reports that the model still blocks credential theft and attacks on third-party systems regardless of access tier.
Capability framing
OpenAI is downplaying the raw capability lift. According to Help Net Security, the company stated that “the initial preview of cyber-permissive models like GPT-5.5-Cyber is not intended to significantly increase cyber capability beyond GPT-5.5.” The Decoder characterizes the model as not smarter than the standard release but less restrictive on security topics.
In benchmarks against Anthropic’s restricted Claude Mythos preview, GPT-5.5 performs roughly on par on cyberattack tasks, according to The Decoder. Citing UK AI Security Institute testing, the outlet reports that GPT-5.5 completed a full attack chain in two of ten attempts, compared with three of ten for Mythos, and is slightly ahead of Mythos on individual expert-level tasks. The Mythos preview itself was at the center of intense industry attention last month after Mozilla credited it with surfacing 271 patched Firefox vulnerabilities, as previously reported.
Vetting and account hardening
Access to GPT-5.5-Cyber requires going through OpenAI’s TAC application, which verifies cybersecurity credentials and intended use, TechCrunch reports. According to The Decoder, starting June 1, 2026, individual users on the highest access tier must enable phishing-resistant authentication.
Altman’s reversal
The optics of the rollout are awkward. TechCrunch reports that Altman previously characterized Anthropic’s gated release of Mythos as “fear-based marketing” before announcing on X that OpenAI itself would push GPT-5.5-Cyber out only “to critical cyber defenders” in the coming days. The same outlet noted that an unauthorized group reportedly gained access to Mythos despite Anthropic’s restrictions, illustrating the practical limits of any vetting regime.
What We Don’t Know
Several details remain unconfirmed in current reporting. OpenAI has not publicly disclosed the size of the GPT-5.5-Cyber preview cohort beyond the TAC scale of “thousands of verified defenders” cited by TechCrunch, nor has the company published a side-by-side benchmark suite against Mythos covering more than the AISI multi-step attack-chain test referenced by The Decoder. The pricing structure for partners outside the named launch list, the timeline for general availability, and any plans to expose GPT-5.5-Cyber capabilities to enterprise customers via OpenAI’s API have not been detailed in the reporting available so far.
Analysis
With both OpenAI and Anthropic now operating cyber-permissive variants under restricted-access programs, the frontier-lab playbook for high-risk capabilities is converging. The pattern — name a small group of large security vendors as launch partners, gate access via identity verification, retain refusals for the most clearly malicious workflows, and publish guarded statements that the new tier is “not intended to significantly increase cyber capability” — looks set to become the default for any future model uplift in offensive security tasks. The competitive question is no longer whether the labs will ship these models, but how quickly the gap between top-tier vetted access and broadly available defensive tooling closes.