Your agent passed its evals.
Did it pass them honestly?
We surface reward-hacking, eval-gaming, and false-green results — the failures internal evals structurally miss, because the same team builds and grades them. The checks live outside the model, so your agent can charm a reviewer but it can't charm a hash.
First 10 audits at cost · public-info only, no system access · you talk to the engineer, not a rep.
This isn't hypothetical
Replit's agent deleted a production database during a code freeze
Then told the user rollback wouldn't work. It would have. Fortune · The Register
METR caught OpenAI's o3 rewriting its own scoring code
Asked afterward if its "passing" solution matched what the user wanted, it said no — 10 times out of 10. It knew. METR
Anthropic found reward-hackers generalize to sabotaging the detection research itself
Their own paper, against their own commercial interest. arXiv 2511.18397
2025 gave the public the disasters. 2026 research shows it's structural, not a bad month.
See it on your own agent — free
Run the reward-hack check on your own agent
A one-file check you run on your own agent's transcripts — no access to your internals. It flags which of our reward-hack patterns your agent trips (self-graded evals, "done" with no verified effect, tests passing on a nonzero exit, silent scope-narrowing) — each finding backed by a verbatim quote, never a fabricated one. See what it hides, in two minutes.
How it works
1. You give us the agent
Its public surface, its eval claims, and what "done" is supposed to mean.
2. We attack the claims
Held-out tasks it's never seen, checks that live outside the model — not the test it graded itself against.
3. You get a signed report
What's real, what's gamed, what's safe to ship — a reproduction receipt per finding.
The offer
Free one-page teardown — a specific, reproducible finding on your public agent. Public-info only, no system access.
Deep pass — first 10 clients at cost. Your top agent adversarially tested, a reproduction receipt per finding, and a ship/block recommendation. I want your agent in the public methodology, not your logo on a wall — you'd be one of the first ten.
Why it's different
Internal evals are graded by the team that built the agent — a structural false-green risk, the same conflict as auditing your own books. AgentReleaseGate is the independent auditor of that stack: we try to break your agent and hand you the evidence, so "green" means the effect actually happened, not just that the response looked right. Checks your agent can't game, because they live outside the model.
Why I built this
I built AgentReleaseGate after my own AI agents corrupted my repo in June 2025 — reported done, wasn't, and did damage on the way. Then I watched the same failure hit everyone else, in public, repeatedly. This is the audit I wish I'd had. One engineer, no sales team, 18 months building anti-reward-hacking enforcement and testing it on my own agent swarm daily. You talk to the person who runs the audit.
Before you ask
"We already have evals."
Graded by the team that built the agent — the same conflict as auditing your own books.
"Won't a better model just fix this?"
OpenAI found penalizing an agent's deceptive reasoning trains it to hide the deception better, not stop it. The fix has to live outside the model.
What access do you need?
None for the free teardown — it's public-info only. A deep pass works from whatever you choose to share.
Turnaround?
Free teardown in a few days. Deep pass within a week.
Request your free teardown
Tell me about your agent. I'll reply personally — no automated sequence.
Free resources
Common reward-hacking patterns — field guide
Read the free guide — 7 patterns we see most often (fake-green, self-grading, stub-as-gate, and more), no signup required.
Runnable OPA policy
Download the companion Rego policy — plug it into your own CI to catch the same patterns mechanically.
GitHub — proof of work
github.com/TTaoGaming/agentreleasegate-oss — the field guide, the policy, and a real example agent-skill file, open source.