cyber.hendryadrian.com
Security Report · arstechnica.com

UK Evaluation Finds Mythos Leads on a Simulated Multistep Cyberattack Test

· Original article ↗

Summary

The UK AI Security Institute found Mythos Preview led models in a simulated 32-step network attack, though it succeeded end to end in only 3 of 10 attempts. The tests do not establish whether it could breach well-defended real systems.

Key points

  • Mythos completed more than 85 percent of the Institute’s Apprentice-level Capture the Flag tasks, a result comparable to recent competing models.
  • In the 32-step “The Last Ones” test, Mythos was the first model to complete the simulated data-extraction attack end to end, succeeding in 3 of 10 attempts.
  • Mythos averaged 22 of 32 infiltration steps on “The Last Ones,” compared with 16 steps for Claude 4.6.
  • The Institute says the results suggest Mythos could autonomously attack small, weakly defended, vulnerable enterprise systems after gaining network access.
  • The simulated ranges lack active defenders and may include vulnerabilities absent from real systems; the tests also do not account for detection stopping an attack.
  • Mythos struggled with the more difficult “Cooling Tower” test, which simulates an attempt to disrupt power-plant control software.

Article Details

Publisher
AI Security Institute (AISI)
Scope
Evaluation of Mythos Preview’s cyberattack capabilities using Capture the Flag challenges and simulated attack ranges, including The Last Ones and Cooling Tower.
Sample Size
10 Mythos Preview attempts on The Last Ones (TLO).
Key Statistics
  • Mythos Preview completed more than 85% of Apprentice-level Capture the Flag tasks.
  • Mythos Preview completed TLO from start to finish in 3 of 10 attempts.
  • The average Mythos Preview TLO run completed 22 of 32 infiltration steps; Claude 4.6 averaged 16 steps.
  • AISI’s tests imposed a 100 million token inference budget.
Recommendations
  • System designers should use AI models to help harden their defenses.

Vendors

AnthropicLast week, Anthropic announced it was restricting the initial release of its Mythos Preview model to “a limited group of critical industry partners,” giving them time to prepare for a model that it said is “strikingly

Products

Claude 4.622 of the 32 required infiltration steps, significantly higher than the 16-step average achieved by Claude 4.6.Claude Mythos PreviewLast week, Anthropic announced it was restricting the initial release of its Mythos Preview model to “a limited group of critical industry partners,” giving them time to prepare for a model that it said is “strikinglyCodex 5.3mark for AISI’s CTF tests, recent competing models like GPT-5.4 and Anthropic’s own Opus 4.6 and Codex 5.3 showed comparable results (within 5 to 10 percent accuracy) across multiple CTF difficulty levels inGPT-3.5 TurboAISI has been putting various AI models through specially designed Capture the Flag challenges since early 2023, when GPT-3.5 Turbo struggled to complete any of the group’s relatively low-level “Apprentice” tasks.GPT-5.4-CyberWhile that’s technically a high-water mark for AISI’s CTF tests, recent competing models like GPT-5.4 and Anthropic’s own Opus 4.6 and Codex 5.3 showed comparable results (within 5 to 10 percent accuracy) acrossOpus 4.6a high-water mark for AISI’s CTF tests, recent competing models like GPT-5.4 and Anthropic’s own Opus 4.6 and Codex 5.3 showed comparable results (within 5 to 10 percent accuracy) across multiple CTF difficulty

Countries

United Kingdomthem time to prepare for a model that it said is “strikingly capable at computer security tasks.” Now, the UK government’s AI Security Institute (AISI) has published an initial evaluation of the model’s cyberattack

Related Articles