Security Researchers Uncover “Option-Channel Attacks” Targeting Typed Decision Models
An October 3 arXiv paper identifies a new adversarial attack surface unique to decision models, showing how option wording and ordering can steer fixed-schema probabilities.
While TypeSafe AI’s launch emphasized that System One models eliminate free-form hallucinations and schema parse failures, a security paper posted to arXiv on October 3 (arXiv:2610.12292) demonstrated that typed decision interfaces introduce their own distinct adversarial attack surface: “Option-Channel Attacks.”
Zero schema errors does not mean zero prompt-injection risk
Because decision models are increasingly deployed as LLM guardrails, tool-approval gates, and spam or fraud filters, attackers have strong incentives to flip a single boolean probability without breaking JSON formatting. The researchers examined how untrusted user text inside the state interacts with the fixed option labels and criteria passed to the model, showing that targeted phrasing can skew option probability mass even when the output remains strictly type-safe.
Hardening Jev-based guardrails
The study arrived alongside related late-September arXiv work examining whether decision models follow option names instead of underlying instructions. For security engineers deploying Jev or rival decision models in front of coding agents and enterprise tools, the takeaway is pragmatic: keep untrusted content cleanly delimited in state fields, test neutral option naming, randomize or audit option order sensitivity, and pair high-stakes actions with deterministic policy checks rather than relying on any single model score.