New Technique Medium TD-2026-0130 For AI Eng, AppSec, CISO AI-assisted
SkillCFI: control-flow integrity enforcement for LLM agent skills (arXiv 2611.24185)
The paper presents SkillCFI, which applies control-flow integrity to procedural agent skills by compiling each skill into agent instructions and an executable contract. A deterministic runtime enforcer admits tool requests only on trusted evidence and refuses unmediated channels such as inline scripts. In 770 agent sessions, the authors report blocking all 240 violations that a permission allowlist admitted.
Why it matters
Permission allowlists alone can admit workflow violations, such as overwriting a file before validation, when the operations are ones the skill legitimately requires. Organizations running skill-based agents may face deviating or prompt-injected models that skip required checks without tripping simple allowlists.
Do this
- Inventory agent skills and identify steps where validation or approval must occur before a tool action.
- Where enforcement is available, require tool calls to be gated on verifiable evidence rather than model claims that a check passed.
- Block or review inline script execution paths that bypass mediated tool channels.
- Evaluate whether allowlist-only controls are sufficient for multi-step agent workflows, and test ordering violations explicitly.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Many violations, such as overwriting a file before validation, use only operations the skill requires, so permission allowlists admit them.”
“In 770 agent sessions, SkillCFI blocks before dispatch all 240 such violations, which an allowlist admits, and normal tasks complete nearly as often as without it.”
Sources
Affected / related
Control-flow integrity for agent skills, SkillCFI
Published 2026-11-18 10:26 UTC · updated 2026-11-19 04:35 UTC · AI-assisted summary, checked against quoted sources · How we score