OpenAI Releases GPT-5.5-Cyber: Battles Anthropic Mythos

On the early hours of April 30, OpenAI confirmed the launch of GPT-5.5-Cyber via a post by Sam Altman on X—a cutting-edge model designed specifically for cybersecurity applications. It will begin deploying to defenders of critical infrastructure within “the next few days.” This move directly corresponds to the cybersecurity model Mythos released by Anthropic in early April, pushing the cyber AI battlefield into a dual-polar OpenAI vs Anthropic setup.

Deployment strategy: government-led trusted access mechanisms

In his announcement, Sam Altman stressed that the deployment of GPT-5.5-Cyber will not go down an open sales route. Instead, it will “work with the entire ecosystem and government agencies to build trusted access mechanisms.” The priority targets are explicit: critical cyber defenders, meaning national-level information security agencies, operators of critical infrastructure (power, water, finance), and government-designated partner organizations. During the pre-release stage, OpenAI has already collected real-world usage scenario feedback from nearly 200 early cooperation partners and completed red-team testing under the OpenAI Preparedness Framework.

This strategy is clearly contrasted with the policy-and-business conflict Anthropic faced when expanding openness for the Mythos model. Anthropic introduced Mythos in early April, and currently about 40 trusted partners have access. When the proposal was to expand to 70 additional enterprises, the White House objected, citing dissemination risks and pressure on government compute resources. In OpenAI’s announcement, the deliberate emphasis on “collaboration with government” and a “trusted access” framework reads like a differentiated positioning from Anthropic’s path—choosing access targets based on a “government gate” rather than “company approval.”

Capability assessment: Preparedness Framework rated High (but not yet critical)

In the system cards and deployment safety page, OpenAI explains that GPT-5.5-Cyber in the cybersecurity domain is rated “High” capability, still below the top tier “critical.” This rating is a specific judgment under OpenAI’s internal safety tiering system. It indicates the model can significantly help cybersecurity personnel improve detection and defense capabilities, but OpenAI does not believe it has reached the “unregulated = constituting a systematic threat” critical level.

While rated High, OpenAI also strengthened corresponding safety protections. The cybersecurity safeguards released this time are stricter than the previously released GPT-5.5 general version. This reflects that the model’s attack capabilities in the cybersecurity domain (not just defense) need access controls to match. Community reports show that on public cybersecurity benchmarks like CyberGym, GPT-5.5-Cyber’s scores are higher than Anthropic Opus 4.7 (about 81.8% versus Opus’s lower level). However, OpenAI’s own official materials have not yet published this specific number, only describing it as a “frontier model.”

Cyber AI dual-polar pattern taking shape

The timing of this release is also worth noting. Anthropic Mythos appeared in early April, and by late April reports surfaced that a plan to expand access was blocked by the White House. OpenAI, on April 30, announced GPT-5.5-Cyber—precisely filling the market gap of what the government wants but Anthropic is blocked from providing. The strategy differences between the two companies in cyber AI are visible in particular:

For Anthropic Mythos / OpenAI GPT-5.5-Cyber Release time: early April 2026 / April 30, 2026 (in deployment) Current access scale: about 40 trusted partners / ~200 early testing collaboration partners Expansion strategy: company vetting partners; planned to expand to +70 by end of April blocked by the White House / work with government to build trusted access mechanisms Capability assessment: company claims it has found “thousands of highly severe vulnerabilities” / Preparedness Framework High (not yet critical)

For readers, points to watch next include: (a) the specific GPT-5.5-Cyber deployment list within “the next few days” and the names of government agencies involved, (b) whether cybersecurity authorities such as the U.S. CISA—previously not granted Mythos access—will prioritize inclusion on the OpenAI cooperation list, © whether negotiations between Anthropic Mythos and the White House will change due to OpenAI’s entry, and (d) whether the two companies will form different camps when setting cybersecurity standards.

This article: OpenAI releases GPT-5.5-Cyber to take on Anthropic Mythos — first appeared at 鏈新聞 ABMedia.