Daily Hypernovelty Lead · AI systems · September 28, 2026
The Risk Appeared Between the Skills
A new study finds that skills approved one by one can produce harmful behavior in sequence. The test was constructed, not an estimate of real-world incidence; it exposes an assurance gap at the handoff.

A paper submitted to arXiv on September 24 examines a blind spot in skill-based AI systems: several skills can look acceptable when reviewed one at a time, then produce a harmful result when they operate in sequence. The authors call this a “skill cascading attack.” The arXiv record says the paper was accepted to NeurIPS 2026, and the study tests the problem across OpenClaw, Claude Code, and Codex. It matters beyond security research because companies are assembling AI workflows from reusable skills, connectors, and agents, often with separate checks for each component. In the paper’s constructed benchmark, safe-looking parts did not reliably produce a safe assembled workflow.
The researchers built 213 validated test cases spanning ten domains, seven harmful objectives, and three patterns of coordination. Each case used an average of 3.25 skills. The cases were tested across eight model backbones, with three runs per configuration. Three expert annotators admitted only cases they unanimously judged valid, and the reported inter-annotator agreement was high.
Across this benchmark, the cascades achieved an average attack success rate of 89.4%, ranging from 76.5% to 97.1% across the tested model backbones. Those figures describe a deliberately constructed stress test. They do not establish how often these attacks occur in deployed systems.
The mechanism is the useful part. One skill can write information into shared context, another can interpret it, and a later skill can act on it. Reviewers who inspect each skill separately may never see the behavior created by the handoff. The dangerous unit is the sequence.
That helps explain why component scanning struggled in the experiments. The paper reports that modified skills cleared individual approval checks more than 86% of the time. Joint scanners still approved most malicious combinations, including 81.7% to 83.0% of the modified stealth variants. Four runtime defenses were also bypassed at rates between 82.5% and 91.8% in the benchmark. Removing a single skill from a three-skill cascade reduced the remaining attack success rate to between 13% and 18%. The behavior depended on the composition.
But this is still early evidence. The work used sandboxed systems, constructed cases, and model-mediated judgments. It does not measure real-world co-installation rates or prove that current enterprise deployments are widely exposed. The proposed behavior-composition scanner reduced cascade stealth by more than 30 percentage points for one class of cascades, while producing smaller gains for other patterns. The defense is a proof of concept, and the hardest interaction problems remain open.
Verification bottleneck
A per-skill pass can become false reassurance when the workflow is the actual decision-making unit. Teams need to know what each component may do, what information crosses the handoffs, and what the assembled system produces under realistic sequences. Those are different questions, and a scanner answering the first cannot silently stand in for tests of the other two.
The practical requirement is to reconstruct the whole path from input to consequence. That means recording skill order, shared-state changes, tool calls, approvals, and final outcomes in a form an independent reviewer can replay. A component inventory remains useful. Its limit should be explicit: it cannot certify interactions it never tests.
Opportunities
The immediate product opportunity is a composition harness for agent workflows. It would run approved skills in realistic combinations, perturb shared context, remove one component at a time, and compare the resulting actions. The aim is to find behaviors that disappear during component-level review and reappear only after assembly. A second layer could provide a workflow bill of materials, recording the name and version of each skill alongside its permissions, context written, context read, downstream consumers, approval points, and rollback ownership.
That combination would give operators a map of where seemingly ordinary outputs become consequential inputs, plus a way to test the map against actual behavior. Approval could then attach to the assembled effect, with component scans treated as one source of evidence rather than the verdict. As AI systems become more modular, assurance has to follow behavior across the seams.