New Technique Medium TD-2026-0027 For AI Eng, AppSec, Red Team, CISO AI-assisted
Indirect prompt injection: task-tampering attacks evade objective-focused defenses
The arXiv paper introduces a 90-case benchmark across three tool-use environments in which planted content tampers with decisions a benign task requires, rather than replacing the task. Without a defense, attacks targeting decision rules, available actions, and state had higher mean success than objective-targeting attacks across five LLMs. A detection-and-removal defense and a tool-policy defense both left substantial attack success on non-objective targets.
Why it matters
Defenses tested only against instruction-override or objective-hijacking prompt injection may overstate their coverage. An attacker who plants content that skews values the agent reads during execution can still alter tool-using agent outcomes, such as inflating a refund.
Do this
- Extend agent prompt-injection test suites to cover decision rules, available actions, observation interpretation, and state beliefs, not only objective hijacking.
- Do not rely on a tool-policy defense alone; validate values that tools return or that agents read from external documents against authoritative sources before they drive decisions.
- Enforce authorization checks so that content from a given source can only influence decisions that source is permitted to determine.
- Review the benchmark code and data at the published repository to reproduce results against your own agent stack before changing controls.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Without a defense, attacks on three of the five other targets (decision rules, available actions, and state) have higher mean success than objective attacks across five LLMs.”
“A detection-and-removal defense reduces objective-targeting attacks to 2.4% success but leaves observation- and state-targeting attacks at 9.5% and 12.1%”
“A tool-policy defense still allows 15.4% attack success, close to 17.0% without a defense”
Sources
Affected / related
Indirect prompt injection, Task-tampering attack
Published 2026-11-05 08:25 UTC · updated 2026-11-08 17:51 UTC · AI-assisted summary, checked against quoted sources · How we score