NaitikVora/mas-injection-testbed

MAS injection testbed

V2 continuation

The live phase summary tracks the sequential v2 study and its $280 lifetime gross Modal cap. CPU Phases 0–1 are complete: all 12 original CSVs recompute identically across 79,345 cells, and the training corpus contains 36,864 operationally labeled utterances with 16 joint holdout folds. The missing prior audit has been reconstructed as a new document.

Gated access to all pinned models has been verified under the configured Modal secret without exposing its value. Phase 2 sequential detector fitting runs under durable budget reservations. The complete Phase 3–6 manifests, workers, offline analyses, and full mock checks are prepared; no result from an unaccepted GPU phase is claimed. The v2 literature reconstruction and deviations log document source adaptations and limits.

The v2 reproduction guide records the exact offline and Modal commands. After fetching the published v2 raw episode chunks and installing the pinned Python/R analysis dependencies, reproduce each accepted phase without model calls using python -m v2.offline phase3 --root . (or phase4, phase5, phase6). This validates exact manifest coverage and regenerates its CSVs, figures, and report. Phase 3 additionally runs the prespecified glmmTMB beta-binomial model. The original 768-episode reproduction command remains below.

A reproducible, exploratory study of wrong-answer propagation in multi-agent LLM networks under repeated exposure to one compromised tool document. The completed Qwen2.5-7B-Instruct dataset contains 768 unique, validated production episodes: four synthetic tasks × three document-order seeds × four graph topologies × four defenses × four source conditions. Six agents interact for eight synchronous rounds; AGENT-2 receives the compromised document each round.

The conditions are clean, corruption only, base injection, and enhanced injection. The corruption-only control separates ordinary misinformation spread from the added effect of injected instructions. The star graph has a hub plus a leaf ring; hub is a pure hub. All nodes are homogeneous research agents, not specialized orchestrators.

Findings

Across graphs, no-defense final non-entry wrong-value matching was 20.0% under corruption alone and 67.1% under each injected payload. The bundled trustguard arm reduced injected-condition matching to 3.3% (base) and 5.0% (enhanced). All 36 exploratory topology-by-defense interaction intervals include zero, so the study does not resolve whether topology modifies defense effects. Trustguard's detector also flagged 122 clean utterances that did not mention the target; its low attack rate must be read alongside that measurement and masking audit.

The adapted ARIA-4 judge disagreed with the deterministic lexical primary outcome on 659 of 3,840 non-entry final answers. It is a secondary, same-model assessment rather than independent ground truth. The full research report gives matched contrasts, 10,000-resample exploratory intervals, propagation curves, clean utility, detector and judge audits, interpretation, and limitations. Results apply to these four synthetic tasks and one fixed model.

Artifacts

• Frozen protocol, execution deviations, and related-work audit
• Raw production JSONL, episode/source manifest, and runtime package/server record
• Conservative budget ledger and validation record with raw SHA-256
• Analysis tables and figures, including all detector decisions and all judge disagreements
• Separate clean pilot and excluded startup attempts

Production contains 46,960 model-call records, including prompts, raw responses, token usage, timings, and retry status. Each episode retains assigned documents, prompt hashes, delivered neighbor messages, eight-round agent traces, trustguard decisions, judge responses, and computed metrics. Raw production JSONL is stored with Git LFS. Mock outputs are mechanics checks and are excluded from empirical findings.

Reproduce the analysis

From the repository root, after downloading Git LFS files:

git lfs pull
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements-analysis.txt
python -m pytest -q
python analysis/analyze.py --input results/production/episodes.jsonl --out results/analysis

The offline command validates exact 768-ID coverage and raw scoring records, recomputes every table and figure, and rebuilds the report without a model call. Eighteen regression tests and a separate 768-configuration mock sweep passed. To repeat the mock mechanics check, run python -m mas_testbed.sweep --stage mock and then python analysis/analyze.py --input results/mock/episodes.jsonl --out results/mock/analysis --allow-mock.

Reproduce the GPU run

An authenticated Modal workspace, access to the pinned model, and a huggingface-token secret for publication are required. Review PROTOCOL.md and the $280 cap before launching a new run; the archived study's conservative ledger totaled $22.70 including an $8 overhead reserve, which is not a Modal invoice. The pilot and production directories are intentionally separate.

pip install modal==1.2.1
modal run modal_io.py --action verify
modal run --detach modal_app.py --action run --stage pilot
# Validate the pilot and its cost projection before production.
modal deploy modal_app.py
python -c 'import modal; f=modal.Function.from_name("mas-injection-study-v2", "run_a10"); print(f.spawn("production", None).object_id)'
python monitor_production.py fc-PASTE-THE-PRINTED-ID

The driver resumes by deterministic episode ID, rejects an incompatible manifest, checkpoints every completed episode to a persistent Modal volume, and reserves a conservative bounded worker cost before starting. The monitor submits another worker only after a normal chunk exit. vLLM and the driver communicate over localhost inside one A10 GPU function; no public inference endpoint is deployed. DEVIATIONS.md records the canceled initial ephemeral invocation, which left 90 valid checkpoints and did not alter the frozen experimental code.

Primary ASR is a wrong-value substring match in final non-entry answers, so it counts mentions and can miss equivalent spellings. Adapted ARIA, APV, and ERS are explicitly different from their source-paper definitions; adapted ERS uses matched clean utility and fixed-range safety normalization. The detector and judge know the target wrong value and share the agent backbone. The payloads are static, seeds vary document order only, and the uncertainty intervals do not establish broad model or task generalization.