Detecting compromised AI coding agents with Jev and Gryph

None. The fetched page is the title plus site chrome with no comments, and the entry notes the thread itself carries no measured numbers beyond the linked post. The figures live in the vendor's write-up, not here.

They come from the linked write-up: an attack set of 14 planted cases that the checks caught, and a pass count of 5,225 out of 5,398. The entry sends you there for the denominators and definitions, which the thread does not supply.

The framing is narrow policy checks layered on top of telemetry, not a complete verdict. The thread does not document what happens when those checks disagree with each other, so you would need the linked article to know how the signal is meant to be gated.

An evaluation write-up from a security vendor, posted as a bare link with no discussion to react to. Read it as one team's test of their own tooling, and check whether the pass-rate denominator resembles your own traffic before drawing conclusions.

Gryph is SafeDep's telemetry for coding agents, and the write-up layers narrow Jev yes/no policy checks on top of it. The entry records the attack and pass figures as coming from the article itself, so the thread is a pointer to a vendor's evaluation of their own tooling rather than a discussion about it.

No. The fetched page is the title plus site chrome with no comments, and the entry notes the thread carries no measured numbers beyond the linked post. There is no discussion of false positives, threshold choice, or what happens when telemetry and the policy checks disagree.

Posted by kunalsin9h with 2 points on 28 September 2026, and the fetched page is the title plus site chrome with no comments. Like several thin threads here, it is a bare link to a vendor write-up with no discussion attached, so nothing beyond the article itself is auditable from the thread.