Use case · Agent security
Hardening agents takes a loop, not a test set.
Agents fail through tool calls, retrieval corpora, and long-horizon plans — surfaces a prompt-only red-team can't see. MEGA fuses static OWASP review with a dual-axis Red Team / Blue Team simulation, then iterates surgical fixes until both defense and over-refusal land inside budget.
Why agents are different
Three failure modes a prompt-only test set never sees.
The optimization loop
Two surfaces, one priority queue. Three exits.
Static review reads source; dynamic probes never do. Findings merge by severity × failure category, then every fix is Pareto-checked on both DSR (defense) and FRR (over-refusal) before it earns its keep.
Two surfaces. One priority queue.
The static review reads source; the dynamic Red Team never does. Both feed independent findings into the hardening loop, merged by severity × failure category.
Pareto every iteration. Three exits.
Every fix earns its keep on both axes — defense up AND over-refusal within budget — or it auto-reverts. The loop ends when there’s nothing surgical left, or the architecture itself needs to change.
Defense leaderboard
Smaller models, hardened, outperform raw frontier.
Validation-set DSR across frontier and small models after the hardening loop converges.
Average defense success rate
Small models with MEGA optimize beat the full frontier average.
Frontier average
Mean across all 4 frontier models
0.000
- Anthropicclaude-opus-4.7·0.910
- OpenAIgpt-5.5·0.830
- Googlegemini-3.1-pro-preview·0.680
- xAIgrok-4.20-0309-reasoning·0.530
Small + MEGA optimize average
Mean across all 4 small models with mega-security:prompt-optimize
0.000
- Googlegemini-3.1-flash-lite-preview·1.000
- xAIgrok-4.1-fast-non-reasoning·0.990
- OpenAIgpt-5.4-mini·0.950
- Anthropicclaude-haiku-4.5·0.910