cyber.hendryadrian.com
Security Report · cyberscoop.com

Claude Mythos and GPT-5.5 Outpace Expected Autonomous Cyber Capability Growth

· Original article ↗

Summary

AISI testing found Claude Mythos Preview and GPT-5.5 exceeding projected autonomous cyber-task growth. Palo Alto Networks also reported rapid AI-assisted vulnerability discovery, while AISI cautioned that its estimates have limitations.

Key points

  • AISI said Claude Mythos Preview and GPT-5.5 exceeded the capability-growth trends it had tracked; whether this marks a lasting change in trajectory remains unclear.
  • Claude Mythos Preview completed both AISI cyber ranges: it solved “The Last Ones” in 6 of 10 attempts and “Cooling Tower” in 3 of 10.
  • GPT-5.5 solved “The Last Ones” in 3 of 10 AISI attempts.
  • Palo Alto Networks said its AI model scanning identified 75 issues covered by 26 CVEs across more than 130 products.
  • Palo Alto Networks said important vulnerabilities in its SaaS products had been patched and patches were available for customer-operated products.
  • AISI noted that its estimates use relatively few models and that the hardest tasks have the least human-comparison data; it is developing more demanding evaluations, including active defenses.
  • Palo Alto Networks urged organizations to prioritize vulnerability remediation, attack-surface reduction, broad detection coverage, and faster security response.

Article Details

Publisher
CyberScoop
Scope
AISI evaluations of autonomous cyber tasks and Palo Alto Networks testing of AI model vulnerability scanning.
Sample Size
Not disclosed; AISI described its model count as relatively small.
Key Statistics
  • AISI estimated earlier this year that the 80% reliability cyber time horizon had been doubling approximately every five months, compared with an estimated eight months in November 2025.
  • In AISI cyber-range attempts, Claude Mythos Preview completed “The Last Ones” 6 of 10 times and “Cooling Tower” 3 of 10 times; GPT-5.5 completed “The Last Ones” 3 of 10 times.
  • Palo Alto Networks reported that AI model scanning across more than 130 products identified 26 CVEs representing 75 issues; its typical monthly volume was fewer than five CVEs.
  • Separate METR research estimated a roughly four-month doubling time for AI software-task capability since late 2024.
Recommendations
  • Find and fix vulnerabilities in code and applications before attackers do.
  • Shrink the attack surface and use AI to identify security misconfigurations.
  • Deploy detection and response tools across systems, using machine learning to detect threats in real time.
  • Build security operations capable of responding in minutes.

Vendors

AnthropicAnthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 have pushed autonomous cybersecurity capabilities beyond the rapid growth trend tracked by the UK’s AI Security Institute, with both models completing advancedOpenAIAnthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 have pushed autonomous cybersecurity capabilities beyond the rapid growth trend tracked by the UK’s AI Security Institute, with both models completing advancedPalo Alto NetworksPalo Alto Networks also found these models can rapidly turn vulnerabilities into exploit paths, prompting urgent enterprise security priorities as AI-driven attacks accelerate.

Products

Claude Mythos PreviewAnthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 have pushed autonomous cybersecurity capabilities beyond the rapid growth trend tracked by the UK’s AI Security Institute, with both models completing advancedClaude Opus 4.7Claude Mythos in April as a launch partner for Anthropic’s Project Glasswing, and has since tested Claude Opus 4.7 and OpenAI’s GPT-5.5-Cyber as part of OpenAI‘s Trusted Access for Cyber program.GPT-5.5Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 have pushed autonomous cybersecurity capabilities beyond the rapid growth trend tracked by the UK’s AI Security Institute, with both models completing advancedGPT-5.5-Cybera launch partner for Anthropic’s Project Glasswing, and has since tested Claude Opus 4.7 and OpenAI’s GPT-5.5-Cyber as part of OpenAI‘s Trusted Access for Cyber program.

Industries

Artificial intelligenceTwo of the most advanced artificial intelligence models — Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 — have significantly surpassed the already-accelerating pace at which AI systems are completing autonomousCybersecurityAnthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 have pushed autonomous cybersecurity capabilities beyond the rapid growth trend tracked by the UK’s AI Security Institute, with both models completing advanced

Related Articles