Taint is the independent security evaluation layer for AI agents — we measure how secure an agent actually is, before it ships and continuously after, and publish it. We're building this now.
The dangerous failures don't live in a single prompt — an agent reads poisoned input, drifts off-task, escalates, and exfiltrates, each step plausible alone. Existing tools check one prompt or one tool call and miss the sequence. Owl Watch scores the whole run.
Any agent from the marketplaces, or your own, connected through our harness.
Adaptive adversarial attacks probe the full trajectory — offline in CI/CD, or online against live agents.
A security score, reproducible failing traces, and drift tracking that updates as the agent changes.
Independent scores for marketplace agents — the neutral reference no guardrail vendor can credibly give you. This is what launches first.
Your agents, under continuous evaluation — before they ship and while they run.
Anyone can claim a high catch rate. What decides whether a control survives in production is its false-interrupt rate — how often it breaks legitimate work. We're building every score to report it, not just catch rate — so it's a real number you can hold us to, not a marketing line.
Nothing's live yet — we're building the leaderboard and the eval platform now. Leave your email and we'll bring you in as it opens, starting with the free leaderboard.