New Technique Medium TD-2026-0029 For AppSec, AI Eng, Red Team, CISO AI-assisted
SPILLWAY: deterministic target admission defense against prompt injection in tool-using agents
The arXiv preprint proposes SPILLWAY, a runtime defense for tool-using agents. It admits an action's target only if the target came from a read-only tool's typed field or matches a fixed delegation rule, and it removes instruction-like text with fixed rules before the model reads it. The authors report that no standard templated injection redirected an action to an attacker's target across three backbones and seven task suites on one backbone.
Why it matters
Agents that act on email, files, and payments can be steered by instructions hidden in web pages or shared documents to send data or money to an attacker. A deterministic check on action targets does not depend on a language model that may read the attacker's text.
Do this
- Identify which agent tools can send data or move money, and list the target fields (recipient, account, destination) each one accepts.
- Restrict action targets to values taken from the user's task or from typed fields returned by read-only tools; block targets that appear only in free text.
- Log every refused action and every span removed by any input sanitizer, so that blocked injections can be reviewed.
- Test the agent against your own templated injection suite and a forged chat-template case before relying on any vendor-reported results; this is a single preprint with self-reported metrics.
Signals
- NoCISA KEV
- 22/100Priority
- 50%Confidence
Evidence
“Where an action goes can instead be read from the tools' typed returns: an email's sender or a payment's account arrives in a typed field, while injected instructions arrive as free text.”
“no standard templated injection redirects an action to an attacker's target under SPILLWAY, at the highest observed mean utility and composite scores”
“mean in-scope attack success under a released reinforcement-learning attacker is .007”
Sources
Affected / related
SPILLWAY, prompt injection
Published 2026-11-04 23:18 UTC · updated 2026-11-09 13:50 UTC · AI-assisted summary, checked against quoted sources · How we score