Work

Per-Attack Blocking Hides Catalog Vulnerability: A Weakest-Link Analysis of Tool Manifests

Year: 2026
Type: preprint
Abstract: Runtime controls for tool-using agents depend on manifests that state what each tool may read and where its output may go. Evaluations of these controls, including ChainCaps, report per-attack blocking: the share of attacks a set of manifests stops. We show that this is the wrong unit. A catalog is only as safe as its weakest source policy: if M independently authored policies are each wrong with probability p, the chance that the catalog contains an accepted exfiltration path is 1 - (1 - p)^M, while per-attack blocking stays at 1 - p. Three lines of evidence agree. In synthetic catalogs run through two enforcement engines, ChainCaps and OpenAPPA, per-attack blocking holds at 92% while catalog vulnerability rises from 22% at 10 tools to 99% at 200, and composition provides no protective effect. A fifteen-model authoring pilot shows that real errors cluster by author, so a 200-tool catalog behaves like about 5 independent decisions rather than 60 (95% interval 3.4 to 9.2). On 114 real MCP tools with expert-written rules, the models' own manifests block 89% of attacks yet leave 9 of 15 catalogs open, as the author-effect model predicts and independence does not. Evaluations should report catalog vulnerability, the sink coverage of their tests, and an effective number of independent decisions alongside per-attack blocking. Accepted at the Third Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD), NeurIPS 2026.
Language: English
Open Access status: green
OpenAlex ID: W7233568469
Cites:
Cited by:
Related to: