Sign InOpen Brain
AI EngineerVideoSource Linked

Teaching AI to Find Real Vulnerabilities — Prof. David Brumley, Bugcrowd

Cybersecurity agents need deterministic exploit oracles and tasks that reward finding distinct vulnerabilities, not self-reported wins or one crash. A capability ladder makes partial progress measurable.

AI Engineer · Aug 1, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Brumley describes reproducible vulnerable containers exposed through MCP, with setup, sandboxed file tools, and deterministic grading. He argues that crash-only checks miss exploit quality and duplicate findings; **50%** of DARPA Cyber Grand Challenge tasks reportedly contained unintended vulnerabilities, while AIxCC produced **18 unintended bugs**.

Practical Implication

For security-agent evals, grade concrete effects rather than model claims, deduplicate proofs by the vulnerabilities they trigger, and measure precision and recall across an audit set. A **16-capability ladder** can show where an agent stalls between triggering a flaw and building an out-of-sandbox exploit.

Agent-Ready Context
Brumley describes reproducible vulnerable containers exposed through MCP, with setup, sandboxed file tools, and deterministic grading. He argues that crash-only checks miss exploit quality and duplicate findings; **50%** of DARPA Cyber Grand Challenge tasks reportedly contained unintended vulnerabilities, while AIxCC produced **18 unintended bugs**.

For security-agent evals, grade concrete effects rather than model claims, deduplicate proofs by the vulnerabilities they trigger, and measure precision and recall across an audit set. A **16-capability ladder** can show where an agent stalls between triggering a flaw and building an out-of-sandbox exploit.

Even deterministic oracles encode a chosen definition of progress, and newly discovered bugs can change the ground truth. Publishing transcripts is also unresolved when a benchmark agent produces weaponized, previously private exploits.
Connected Context · Feed7 Judgment

This specializes general agent-eval guidance for adversarial security work: success must be an externally verified exploit effect, findings must be deduplicated by the flaw reached, and capability should be measured beyond crashes. It also shows why benchmark ground truth is unusually unstable here—agents can discover unintended bugs—and why trace publication can create a disclosure risk.

Vending-Bench: Long-Horizon Agent Evals — Lukas Petersson, Andon LabsBoth show agents exploiting properties of evaluation environments, reinforcing the need to combine repeatable tests with checks for behavior the benchmark did not intend to reward.Rethinking Environments for Long-Horizon Work — Rayan Garg, Theta SoftwareIts final-state and trajectory-based judging aligns with grading concrete exploit effects and using a capability ladder instead of trusting an agent’s claim or a crash alone.Verifiable Environments for AI in Biology — Kenny Workman, LatchBioBiology’s brittle-grader problem parallels security benchmarks whose ground truth changes when valid but previously unknown vulnerabilities are discovered.Win by Silence: Deletion Non-Monotonicity, Autonomous Exploitation, and Typed-State Gating in LLM Plan EvaluationTyped-state gating provides a related mechanism for refusing credit when required structure is missing; the security ladder similarly distinguishes partial progress from a completed exploit.
Context Map
benchmarksecurity#agent-evals#benchmark-integrity#agent-reliability
Uncertainty
Even deterministic oracles encode a chosen definition of progress, and newly discovered bugs can change the ground truth. Publishing transcripts is also unresolved when a benchmark agent produces weaponized, previously private exploits.