SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
SPILLWAY blocked every standard templated attack from redirecting an agent’s action to an attacker’s target in the reported tests.
Its deterministic check admits destinations found in tool output only when they come from typed fields returned by read-only tools or meet a fixed delegation rule. A separate sanitizer strips instruction-like text before the model reads it; a local judge can restore refused calls, but cannot approve targets found in the stripped text. Across three backbones, SPILLWAY also recorded the highest mean utility among the defenses tested under those attacks, with almost no added model calls. A released reinforcement-learning attacker achieved a mean in-scope success rate of 0.005, so the reported protection was not absolute.
HF Daily Papers' note
Its deterministic check admits destinations found in tool output only when they come from typed fields returned by read-only tools or meet a fixed delegation rule. A separate sanitizer strips instruction-like text before the model reads it; a local judge can restore refused calls, but cannot approve targets found in the stripped text. Across three backbones, SPILLWAY also recorded the highest mean utility among the defenses tested under those attacks, with almost no added model calls. A released reinforcement-learning attacker achieved a mean in-scope success rate of 0.005, so the reported protection was not absolute.
HF Daily Papers' note
score 5