Vibeleaderboard
← All Intel
Intel / article

How we built Datadog’s agentic vulnerability hunting system

Source
datadoghq.com
Date

Context

Datadog's approach to noisy AI vulnerability reports is to give its hunting agents a written statement of what each system must never allow, then stop pursuing any finding that cannot show how an attacker would break one of those rules. In a post on its engineering blog, Datadog's security engineering team says that current models are good at spotting suspicious code. The hard part is judging whether it is exploitable when the authentication check or network restriction that would block the attack lives in another team's service, outside the repository the agent is reading.

The written statement is a threat pack: a versioned Markdown file of about 300 lines, plus metadata, covering one service or trust boundary. It records a system summary, review priorities, known compensating controls (protections elsewhere that stop an attack, such as an egress proxy that is the only approved route to the internet), and security invariants, meaning properties that must always hold. One example: code running inside an isolated sandbox must never reach the Kubernetes node, the host filesystem, the container runtime or the cloud metadata service. Agents draft each pack from code, penetration test reports and design documents, then interview the service's experts about what the code cannot show, and a human reviews the pack before any hunt uses it. Datadog says it has built hundreds.

The invariants then act as the filter. A second agent, acting as a judge, accepts a candidate finding only if it names a declared invariant, identifies an entry point an attacker can reach, and explains the path from one to the other. Accepted findings get a proof-of-concept attempt in staging or locally before triage, and rejected ones are kept so reviewers can spot gaps in the pack. Datadog says this has reduced noise and improved report quality, without publishing figures, and that hunts using packs found guest-to-host escapes, where code inside a virtual machine sandbox can reach the machine hosting it, in open source components, including one through virtio-fs in Kata Containers' Rust runtime and one from missing length validation in Dragonball's virtio-blk implementation.

The other half of the system tests the hunters. Agents can return different findings from identical inputs, so Datadog reruns frozen hunts against a benchmark of known vulnerabilities, 43 of them across nine threat packs in its initial August 2026 version, with the answers hidden from the agent and a separate judge matching results by root cause and attack path. A vulnerability counts as stably detected when it turns up in at least 7 of 10 runs, a threshold Datadog calls provisional. That lets the team see whether a new model, prompt or finds known bugs less consistently. Datadog has not published detection rates, so the post documents a method and some concrete finds rather than measured accuracy.

Key takeaways · AI-distilled
  • Datadog's threat packs pair a versioned Markdown file (system summary, security invariants, review priorities, known compensating controls) with a YAML file naming scope, maintainer and code locations, kept near 300 lines to use little of an agent's .
  • Agents draft each pack from code, penetration test reports, risk reviews and design docs, then an interview has them question subject matter experts. Every pack gets human review before a hunt; Datadog says it has built hundreds, prioritized by risk.
  • A judge agent qualifies a candidate finding only if it names a declared invariant, an attacker-reachable entry point and the path between them. Datadog keeps disqualified reports to find gaps in packs and benchmarks the filter so it does not drop real bugs.
  • A Markdown renderer that loaded external images let leak data through an image URL. Datadog recorded a hardened-renderer rule as a UI-wide invariant, and within days hunts surfaced several more confirmed instances of the pattern.
  • To test model or harness changes, Datadog reruns hunts on a frozen benchmark (43 known vulnerabilities across nine packs as of August 2026). A semantic judge matches results, and 'stable detection' means finding a bug in at least 7 of 10 runs, a working threshold.
Terms in this piece · Glossary
  • agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
  • prompt injection — An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.
Read the source www.datadoghq.com
Recommended reads
Comments

Checking sign-in…

Loading comments…