GPT-5.5 Matches Claude Mythos in Cyberattack Simulation Tests

Two Top-Tier Models Demonstrate Alarming Capabilities in Cyberattack Simulations

The latest test results from the UK AI Safety Institute reveal that OpenAI's GPT-5.5 has become the second AI model capable of autonomously completing full cyberattack simulations, nearly matching the performance of Anthropic's Claude Mythos. This finding has once again thrust the safety risks of AI models into the spotlight.

What makes this particularly noteworthy is that GPT-5.5 is already publicly available through ChatGPT and its API, while Claude Mythos remains accessible only to a limited user base. This means an AI model capable of autonomously completing cyberattack simulations has already been widely deployed on a global scale.

Test Details: A Security Wake-Up Call from Simulation to Reality

According to the UK AI Safety Institute, the evaluation employed a full cyberattack simulation environment to test whether AI models could autonomously identify vulnerabilities, formulate attack strategies, and execute complete attack chains. Previously, only Anthropic's Claude Mythos had demonstrated this level of capability, and GPT-5.5's entry marks a rapidly accelerating advancement in frontier AI models' cybersecurity capabilities.

The prevailing industry view had been that while AI models could assist security researchers with vulnerability analysis and penetration testing, they were still a considerable distance from fully autonomously executing complex cyberattacks. However, the test results from GPT-5.5 and Claude Mythos indicate that this gap is closing far faster than expected.

Differences in Availability Create Distinct Risk Dimensions

The disparity in availability between the two models is one of the most critical aspects of this finding. Claude Mythos remains in a limited-access phase, with Anthropic maintaining strict controls over its usage. In contrast, GPT-5.5 has been fully integrated into the ChatGPT product line and made available to developers worldwide via API.

This situation — equal capability but vastly different accessibility — has raised concerns among security researchers. On one hand, broader availability means more security researchers can leverage these capabilities for defensive research. On the other hand, it also significantly increases the risk of potential malicious use.

Industry Impact: New Challenges for AI Safety Governance

These test results pose new challenges for global AI safety governance frameworks. As the world's first national-level AI safety evaluation body, the UK AI Safety Institute's findings carry significant policy implications.

From a technological evolution standpoint, the advancement of AI model capabilities in cybersecurity reveals several clear trends:

• Rapidly lowering capability thresholds: The time gap between only one model possessing this capability and two models achieving parity has been remarkably short
• Commercialization of frontier capabilities: Models with advanced security capabilities are no longer confined to laboratories but are being offered directly to consumers
• Urgent need for upgraded evaluation frameworks: Traditional AI safety benchmarks may be insufficient to capture these emerging risks

Outlook: The Double-Edged Sword of Security Capabilities

The capabilities demonstrated by frontier AI models in the cyberattack domain are fundamentally a double-edged sword. On the positive side, these capabilities can be harnessed for automated security audits, vulnerability discovery, and defense system development, dramatically improving the efficiency of the cybersecurity industry. On the negative side, without effective usage restrictions and regulatory frameworks, the same capabilities could be exploited in real-world cyberattack scenarios.

With GPT-5.5 and Claude Mythos both demonstrating this level of security capability, it is foreseeable that AI safety regulators across nations will accelerate the development of mandatory safety evaluation systems for frontier models. Leading companies such as OpenAI and Anthropic will also face mounting pressure to find a more refined balance between opening up model capabilities and controlling security risks.

The race in AI safety is shifting from "whose model is more powerful" to "whose safety guardrails are more reliable."