The next prompt injection your coding agent reads might be an error report. So Sentry screens for it first, spending $263 a day on Jev to check 10M events. Now the team is testing its own classifier, aiming to match Jev's accuracy for 90% less. Greg Pstrucha explains how.

@grichadev building an in-house classifier to save $230 a day is peak infra behavior

boy i need some beta blockers next time

@grichadev For 2000$ per month? What

@grichadev Stack traces as an attack surface is the kind of thing nobody writes down until it bites someone. Does the classifier sit in front of the agent, or does the agent read the raw event either way?

@grichadev Screening error reports before the agent reads them is smart, and it will always run a step behind. Any text an agent reads can steer it, and a classifier is one more model guessing. What has to hold is the action: a check on every tool call, outside the model, default no.

@grichadev Error reports are a sneaky place for this. I'd want the original event kept intact for debugging, with the suspicious bits clearly marked before the agent reads it.

@grichadev At 10M events a day, matching Jev's accuracy is only half the job. How are you planning to catch drift when attackers start shaping error reports around the in-house classifier?

@grichadev Why not use clef?

@grichadev The attacker's new job is writing an error report that the cheap classifier calls safe and the coding agent still obeys

@grichadev Thought you had clef!?