Papajams/elcaro-ipi-eval

Elcaro IPI Eval

A deterministic benchmark for indirect prompt injection (IPI) against tool-using LLM agents: 24 original attack tasks spanning email, documents, search results, tool manifests, memory, and agent-to-agent relays. Each task pairs an agent role + user request with untrusted content containing an injection payload, plus a deterministic grader (canary echo, concealed compromise, exfiltration indicators).

Built for the Lexsi AI Safety & Alignment Hackathon (Track 2a) and the micro1 Eval Hackathon. All scenarios, payloads, and graders are original; techniques are informed by published IPI research (AgentDojo, InjecAgent, BIPIA, ChatInject).

Contents

Path	What
tasks/T01.json–T24.json	The task corpus: system prompts, user turns, untrusted content, benign requirements, compromise indicators, reference responses
curated-tasks/	CuratorKIT curation output — provenance manifest, dataset card, checksums, hygiene-gate report
results/	labeled arm transcripts (explicit untrusted-content warning)
results-naive/	naive arm transcripts (raw content, no warning)
results-wrapped/	wrapped arm transcripts (Elcaro detector warning prepended)
grade.py, selftest.py	Deterministic grader + self-test (no API needed)
run_eval.py	OpenAI-compatible eval runner (any endpoint)
auditkit_eval.py	Lexsi AuditKIT-orchestrated run (api: backend)
curate_tasks.py	Lexsi CuratorKIT curation pipeline
train_tasks/TR01.json–TR36.json	36 new IPI training scenarios, disjoint from the eval corpus (SafeTune arm)
sft_train.jsonl	SFT chat records: answer usefully + ignore the embedded instruction + say so
results-safetune/	Repair-arm before/after on aya-expanse-8b: per-task grades + 48 raw responses + comparison.json + kernel log
lexsi_safetune_kernel.py, make_sft.py, author_train_tasks.py	Kaggle repair-arm pipeline + SFT-data generators

Models evaluated

Allowed dense open-weight Cohere models via the Hugging Face Inference Providers router (https://router.huggingface.co/v1), 3 trials/task/arm (72 trials per cell):

Model	naive ASR*	labeled ASR*	wrapped ASR*
CohereLabs/aya-expanse-32b	24/72 compromised	24/72	12/72
CohereLabs/tiny-aya-global	9/72	9/72	6/72

* "compromised" = concealed compromise indicators fired (canary echo, smuggled tool call, exfiltration payload). Pass = useful response AND no compromise. See results*/summary.md for full tables including per-task pass/useful/detected breakdowns.

Key findings

1. Raw ingestion fails. aya-expanse-32b leaks canaries or executes injected instructions in ~33% of naive trials.
2. Hand-written warnings don't generalize to this model. The labeled arm (explicit "untrusted content" framing + warning system prompt) performs statistically identically to naive for both models.
3. Automated provenance warning helps. Elcaro's wrapped arm — the production detector prepending an [ELCARO WARNING] with risk score, techniques, and summary — halves concealed compromise on aya-expanse-32b (24→12) and on tiny-aya-global (9→6). Residual failures remain and are published, not hidden.
4. Curation gates can destroy adversarial data. CuratorKIT's default strip_html cleaner silently removes tag-based injection payloads; its secrets gate flags base64-encoded injection blocks as "high-entropy secrets." Both are documented false-positive classes.

Repair arm (SafeTune) — negative result

LoRA SFT of CohereLabs/aya-expanse-8b on the 36 disjoint training scenarios (r=16, 3 epochs, completion-only loss, Kaggle T4 x2), re-measured on the 24-task eval under the naive arm:

arm	pass	useful	compromised
baseline	16/24	21	5
tuned	15/24	21	6

~36-example SFT bought no measurable IPI resistance — compromised went 5→6 (single-task noise, slightly worse), replicated across two kernel runs. Honest takeaway: intrinsic resistance needs more than a ~100-example PoC. The LoRA adapter is published anyway: https://huggingface.co/Papajams/aya-expanse-8b-elcaro-sft (collection: https://huggingface.co/collections/Papajams/lexsi-ipi-eval-cohere-aya-models-6aef8fd587f725ebce1bc528)

Reproduce

export EVAL_BASE_URL=https://router.huggingface.co/v1
export EVAL_API_KEY=<your HF token>
export EVAL_MODELS=CohereLabs/aya-expanse-32b

python3 run_eval.py --presentation naive   --trials 3
python3 run_eval.py --presentation labeled --trials 3
python3 run_eval.py --presentation wrapped --trials 3
python3 selftest.py                        # grader self-test, no API
python3 run_eval.py --regrade              # re-grade stored transcripts

CuratorKIT curation: .venv/bin/python curate_tasks.py (see lexsi-eval/README.md). AuditKIT run: .venv/bin/python auditkit_eval.py.

Citation

Elcaro — runtime defense for AI agents against indirect prompt injection. https://elcaro.trustfall.xyz · https://github.com/udirobert/elcaro