Korea University researchers propose ActionGuard, which checks skill-influenced tool calls immediately before execution against the trusted user request and runtime evidence. In an OpenClaw evaluation using SKILL-INJECT tasks, they report 8.65% overall attack success and 90.38% task success, averaged across reviewer models and injection types, with task success evaluated on a restricted subset. The evaluation uses one framework, one target model and one benchmark; it does not establish performance against attackers adapting to the defense.
It tests authorization at the execution boundary instead of treating skill instructions as permission, while measuring the resulting task-completion cost.
Key facts
As stated in the sources, with where to find them.
- ActionGuard separates the agent’s planning context from an isolated reviewer’s authorization context. The reviewer uses the trusted request, an independently maintained skill profile, tool-call context and local script contents, rather than receiving the raw potentially poisoned skill as execution authority. Invalid or unavailable reviewer decisions deny execution.v1, Sections 4.1–4.3
- Evaluation uses 139 contextual and 180 obvious injection-task pairs in OpenClaw, five reviewer models, and three repetitions per condition. Comparisons include Dynamic Guardian, SkillGuard and no safeguard. Attack success uses all 319 pairs; task success uses 255 pairs (100 contextual and 155 obvious), excluding destructive cases whose benign-task completion cannot be judged. The authors reimplement the two baselines in OpenClaw and use GPT-5.5 as both the target and outcome judge.v1, Sections 5.1–5.3; Tables 2–3
- Table 4 averages across reviewer models and injection types: ActionGuard 8.65% attack success and 90.38% task success; Dynamic Guardian 16.05% and 92.08%; SkillGuard 13.42% and 88.55%; no safeguard 29.05% and 93.46%. These are averaged rates, not individual-model results.v1, Table 4; Section 6.2
- The authors limit generalization to one agent framework, target model and benchmark. Blocking a call that combines benign and unauthorized actions can leave the task incomplete if the agent does not produce a safe alternative. Their abstract and conclusion give slightly different relative reductions against no safeguard; this record uses the absolute Table 4 rates. ActionGuard’s lower attack success is aggregate: SkillGuard is slightly lower on obvious injections (8.48% versus 9.00%), and Dynamic Guardian has higher task success overall.v1, Limitations and Conclusion; Abstract; Tables 3–4
Findings that cite this record
No tracked finding cites this record yet.
Key questions this bears on
- Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.
- Where are deployed AI agents actually being exploited?Exploitation often reaches deployed agents through connectors, credentials, tools and packages.
- Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense lacks operational validation and production oversight of agents has documented gaps.