Defensive benchmark. leakgauge measures whether an agent's defences hold
against prompt injection, so they can be improved — it is not an attack toolkit. Success is
programmatically verified data leakage, never LLM-judged. See
THREAT_MODEL.md.
Final result: no. Ranking these 3 models by verified data leakage gives the same order as ranking by task hijacking (Kendall τ = 1.000 — the lines above are parallel, no crossing). The leakage-vs-hijack reorder does not occur for this roster.
Why, and what it means. On current frontier models these attacks score near the floor — the models barely leak, and hijack-ASR ≈ leakage-ASR. With so little spread there is almost nothing for leakage-vs-hijack scoring to reorder. Demonstrating the reorder needs harder cases that pull the two rates apart (future work). The contribution here is the programmatic leakage-verification methodology and a verified leaderboard — not a positive reorder.
Limitations. qwen-2.5-72b stalled provider-side and is excluded. Anthropic (Sonnet/Haiku) was deferred on cost — agentic runs take many tool-call steps (~$15–60/model) and, being similarly robust, would not flip this null.
Per-model rates (95% bootstrap CI)
model
hijack-ASR
leakage-verified ASR
utility-under-attack
n
openai:gpt-4o
0.08 [0.03, 0.15]
0.08 [0.03, 0.14]
0.66 [0.52, 0.79]
37
openrouter:meta-llama/llama-3.3-70b-instruct
0.02 [0.01, 0.04]
0.02 [0.00, 0.03]
0.43 [0.30, 0.57]
37
openai:gpt-4o-mini
0.00 [0.00, 0.01]
0.00 [0.00, 0.01]
0.78 [0.66, 0.90]
37
Rates are means over cases with a percentile bootstrap 95% CI. The suite is small
(n = 37), so the intervals are wide — shown openly rather than
hidden. The highlighted column, leakage-verified ASR, is the
metric the project is about.