mythos-ai-llm-probe-mode
Build defender-side LLM safety systems using activations-based probes, classifier ensembles, and prompt-injection / jailbreak detection - the Anthropic Safeguards approach
More in Mythos
mythos-behavioral-analysis-mode
Automated transcript analysis of autonomous cyber-agent runs — unique services, exploit/exploration ratio, credential reuse, drift detection, cost-per-milestone
View → modemythos-context-compaction-eval-mode
Implement and audit context compaction for long-horizon agent runs — ~80% trigger, summarization fidelity, KV-cache cost tradeoffs, credential handling
View → modemythos-cyber-eval-disclosure-mode
Responsible disclosure norms for cyber-eval results — methodology without exploit recipes, hash-then-reveal pattern, coordinating with developers and governments
View → modemythos-cyber-range-designer-mode
Design multi-step, milestone-graded cyber ranges for AI agent evaluation, modeled on AISI's "The Last Ones" and "Cooling Tower"
View → modemythos-inference-scaling-eval-mode
Run token-budget sweeps against autonomous cyber agents to measure log-linear inference-time compute scaling, modeled on AISI's 10M→100M findings
View → modemythos-inspect-ai-harness-mode
Build agent evaluations on the UK AISI Inspect AI framework with Sandboxing Toolkit integration, real Task/Solver/Scorer APIs
View →