fix: carry the re-measured numbers, after the benign set grew - #11
Merged
Merged
Conversation
The benign set went from 63 paragraphs to 68 when the tool repository took a contributed source, the promptfoo page on indirect prompt injection. Six requests went out to fill the gap: the five new paragraphs, and one existing row whose text had changed upstream because its source is fetched live rather than pinned. False positives at the annotate line go from 7 of 63 to 8 of 68, and from 10 to 13 at 0.3. The other thresholds are unchanged, and so is every other set. The new flag is a table of attack payloads that lists injection strings verbatim as documentation, which is the case this set exists to measure, so one more here is the set working rather than the screen getting worse. The run date moves with it. The other 1,941 responses are still the ones from the earlier date and every new one came from the same model version, so this is the same run extended rather than a different one. Everything on the site reads these four values, so the home page, the results page, the prompt-injection guide and both markdown twins move together.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The benign set grew, so the published numbers moved with it.
agent-chaperone/agent-chaperone#70added the promptfoo page on indirect prompt injection to the benign corpus, taking it from 63 paragraphs to 68. Re-running sent six requests: the five new paragraphs, plus one existing row from the guardrails cookbook whose text had changed upstream, because that source is fetched from a live URL rather than pinned.Every other set is identical, question for question. The new flag is the promptfoo table of attack payloads, which lists injection strings verbatim as documentation and scored 0.54. That is precisely the case this set exists to measure, so one more false positive is the set doing its job rather than the screen getting worse, and the two highest scoring rows are unchanged.
The run date moves to 2026-09-21 because that is when these numbers were complete. The other 1,941 responses are still the ones recorded on 2026-09-19, and all six new ones came back from
jev-1.13.0, so it is the same run extended rather than a different one.Nothing here is written into a page. All four values live in
src/data/benchmark.ts, so the home page, the results page, the prompt-injection guide and both markdown twins moved together, which was the point of keeping them there.Lint, typecheck, build, format, the spacing guard and the twin check all pass.