ReliaQuest Describes Defensive Testing of Anthropic’s Claude Mythos
Summary
ReliaQuest says it is using Anthropic’s Claude Mythos for defensive testing, reporting results in attack-path discovery, exploit development, attacker-behavior prediction, detection review, and remediation prioritization.
Key points
- ReliaQuest says Mythos identified a cloud attack path involving an obscure API function with hidden access implications.
- In a PHP2Shell rebuild involving four bugs across two CVEs, ReliaQuest says Mythos generated an actionable proof of concept in one day.
- In a 41-stage synthetic macOS infostealer chain, Mythos 5.1 had 39% first-guess accuracy for predicting the next attacker activity, according to ReliaQuest’s tests.
- ReliaQuest says Mythos reviewed part of its detection library in 1.5 hours, compared with 4.5 hours for a baseline frontier model, and used roughly 50% fewer interaction turns.
- The company says it is using Mythos to assess whether controls block attack paths and to prioritize exposures, configuration risks, and remediation.
- ReliaQuest recommends testing defensive controls, auditing telemetry, mapping likely attack paths, and modernizing response processes.
Article Details
- Event Type
- ReliaQuest is testing Anthropic’s Claude Mythos for defensive cybersecurity use under a partnership.
- Impact
- ReliaQuest reports that testing identified a hidden cloud attack path, generated a weaponized proof of concept from exploit theory in one day, and improved prediction of activity in a synthetic macOS infostealer attack chain. In a detection-library review, Mythos completed analysis in 1.5 hours versus 4.5 hours for a baseline frontier model and used roughly 50% fewer interaction turns. These are company-reported test results, not reported impacts from a real-world attack.
Vendors
Products
Claude Haiku 4.5likely next move. Mythos 5.1 achieved the top first-guess accuracy at 39%, followed by Opus 5.5 (34%), Haiku 4.5 (15%), and Sonnet 4.6 (12%).Claude MythosClaude Mythos represents an important step forward in cybersecurity. Through our partnership with Anthropic, ReliaQuest is among the few to have access to Mythos, actively deploying its advanced reasoning capabilitiesClaude Opus 5.5the attacker’s likely next move. Mythos 5.1 achieved the top first-guess accuracy at 39%, followed by Opus 5.5 (34%), Haiku 4.5 (15%), and Sonnet 4.6 (12%).Claude Sonnet 4.65.1 achieved the top first-guess accuracy at 39%, followed by Opus 5.5 (34%), Haiku 4.5 (15%), and Sonnet 4.6 (12%).Mythos 5.1In testing across a 41-stage synthetic macOS infostealer attack chain, Mythos 5.1 led the other models in predicting the attacker’s likely next move. Mythos 5.1 achieved the top first-guess accuracy at 39%, followed by