Essential session storage is required to sign in. Optional analytics remain off unless you allow them.
Design (and, where you can, implement) a DEFENSIVE safety EVALUATION suite for one LLM feature. Build eval sets that measure things like harmful-content refusal, jailbreak/ bypass RESISTANCE (using only your own benign test cases — no sharing of working exploits), prompt-injection robustness for any retrieved content, and false-refusal rate on benign inputs. Define the metrics (e.g. attack-success rate, refusal rate, false-refusal rate), use programmatic checks and/or LLM-as-judge to score, and show how you'd measure robustness BEFORE and AFTER adding a defense (e.g. an output classifier or input validation). Document rules of engagement: everything is tested against a system you own/are authorized to test, with synthetic data.
Covers refusal, bypass resistance, injection robustness, AND false-refusal on benign inputs; sound metrics; programmatic/LLM-judge scoring.
Measures robustness before and after a defense to prove it helps; regression-capable.
Tests only an owned/authorized system with synthetic data; rules of engagement documented; no exploit sharing.
Use this README structure to make the work reviewable.
DevPath fetches a read-only, size-capped snapshot. Submitted code is never executed.
# A safety eval suite for an LLM feature ## Goal Design (and, where you can, implement) a DEFENSIVE safety EVALUATION suite for one LLM feature. Build eval sets that measure things like harmful-content refusal, jailbreak/ bypass RESISTANCE (using only your own benign test cases — no sharing of working exploits), prompt-injection robustness for any retrieved content, and false-refusal rate on benign inputs. Define the metrics (e.g. attack-success rate, refusal rate, false-refusal rate), use programmatic checks and/or LLM-as-judge to score, and show how you'd measure robustness BEFORE and AFTER adding a defense (e.g. an output classifier or input validation). Document rules of engagement: everything is tested against a system you own/are authorized to test, with synthetic data. ## Acceptance criteria - [ ] Eval design: Covers refusal, bypass resistance, injection robustness, AND false-refusal on benign inputs; sound metrics; programmatic/LLM-judge scoring. - [ ] Before/after rigor: Measures robustness before and after a defense to prove it helps; regression-capable. - [ ] Authorized & defensive: Tests only an owned/authorized system with synthetic data; rules of engagement documented; no exploit sharing. ## Verification - [ ] Document installation and run commands. - [ ] Record automated checks and their exact commands. - [ ] Add representative output, screenshots, or a short demo where useful. - [ ] Confirm failure paths and known constraints. ## Decision log ### Decision title - Context: - Choice: - Alternatives considered: - Tradeoffs: ## Evidence - Link each rubric criterion to the file, test, or artifact that demonstrates it. ## Known limitations - List what is intentionally out of scope and what should be improved next.