Sentinel detection metrics
Measured, not claimed — the weekly replay of Sentinel's live detectors over public exploit windows and random blocks: recall, lead time and false-positive rates.
Measured, not claimed
Vendors in this space publish detection rates nobody can check. Athena Sentinel publishes the opposite: every week the exact detector code that runs live is replayed over public exploit windows and over ordinary blocks, and the outcome lands here unedited — including the misses.
The replay uses public incident lists, the free data sources the live tail uses, and the label set as it stood before each incident. A detector that scores well here can still miss the next exploit; one that scores badly is telling you exactly what it does not see.
Latest evaluation
How the replay works
The incident set is every exploit on Sentinel’s EVM chains catalogued in a public, permissively licensed exploit archive since September 2025, joined with public loss estimates. For each one, the blocks around the attack transaction are fetched the same way the live tail fetches them and pushed through the production detectors. Labels first seen in or after the incident month are hidden from that replay, so an exploiter’s own label can never make its incident look detected. A sample of random recent blocks is replayed the same way to measure how often the detectors fire when nothing is happening.
What the columns mean
- Recall — the share of incident windows in which the detector produced a signal strong enough to be alerted (the same bar the Enclave feed uses).
- Lead — attack block minus first signal block. Positive means the detector fired before the exploit transaction (for example on a privileged-control change); “after” means it caught the laundering or the issuer’s reaction, not the attack itself.
- False positives per day — alert-grade signals on the random blocks, scaled to a day of that chain. Issuer freezes are ground truth and count as signal, not noise.
Known limits
The v1 detectors cover issuer freezes, tainted funds on the move and privileged-control changes on a curated set of protocols. They are not exploit detectors: a fresh attacker contract draining an untracked protocol produces no signal in this version, and the table says so. Detector thresholds and the tracked-protocol list are not published. A small fixed subset of these replays runs on every code change and blocks the change if a previously detected case stops firing.
See also Security alerts and the risk model.