AI NewsProductivityAnnouncement

Datadog built an agent that hunts security bugs using 300-line threat packs

Datadog wrote 300-line security briefs for each service, then let AI agents hunt for exploits. It benchmarks against 43 known vulnerabilities.

AI News

Editorial2 min read

LinkedInX
Datadog hero graphic for the agentic vulnerability hunting system post

Image: Datadog

Why it mattersA team building a security agent now has a working design to compare against. The threat pack format is the part worth copying.

Most security agents run on a blank prompt. Datadog published the design of its own agent on 18 November. The authors are Bianca Lankford, Matt Meen and Adrien Peronnet. The piece names the parts of the system: threat packs, discovery agents, judge agents, and a benchmark the team checks progress against. The design runs on Datadog's own code.

The post has a clear claim about the hard part. Finding a candidate bug is not it. Deciding whether that candidate is exploitable in a specific service is. "An agent reviewing one repository may not see the authentication checks or network restrictions" another team enforces, Datadog writes. The threat pack is where that missing context lives.

What a threat pack is

A threat pack is a small document for one service. It pairs a Markdown file with YAML metadata. Datadog says each pack is usually around 300 lines. It holds the system summary. It names the security invariants the service relies on. It lists the review priorities for that codebase. It names the compensating controls already in place. The size is deliberate. 300 lines fits inside a model's context window alongside the code the agent is reviewing.

Datadog has written hundreds of these packs, prioritised by risk. The post calls them "compact, versioned documents". A pack is read by agents and by people. A new team member reads the same document to learn what to worry about.

How the agents work together

The system has three roles. Discovery agents analyse code against a threat pack and propose candidate findings. Judge agents qualify those findings by checking whether an exploit is reachable given the service's business logic and controls. Semantic judges compare new findings against known ones during benchmarking. A hunt is scored against ground truth, not against itself.

Datadog's initial benchmark, dated August 2026, held 43 validated vulnerabilities across nine threat packs. A finding is counted as stable when the system detects it in at least 7 of 10 repeated runs. Datadog uses hidden reference findings during evaluation, so the agents cannot be tuned to the test.

What the agents found

The post names two guest-to-host escape vulnerabilities in Kata Containers. One is in the Rust runtime through virtio-fs. The second is in Dragonball, Kata's virtualisation component, through a missing length check in virtio-blk. Both cross the boundary between a guest and the host, which is the boundary the whole product exists to defend. Datadog did not name the exact severity or the patch dates for either in this post.

Source

How we built Datadog's agentic vulnerability hunting system, Datadog.

SourceDatadog

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project