Information flow control
IFC · information-flow control · data flow tracking
Tracking where each piece of data came from and who may see it, then blocking any step that would send it somewhere it should not go.
In our briefings
The claim each of the newest briefings on Information flow control would have you leave with.
The authors argue that least privilege cannot reach an agent the way it reached phone apps, where a developer declares the permissions up front. "In agentic systems, there is no app developer and there is no app," they write, only a natural-language task that can change, so the agent's privilege "needs to be predicted from the task description." Their example is a browser agent asked to comment on GitLab issue 7745. It needs that issue page and write access to its comment box, and nothing else in GitLab or the user's browser session. A follow-up request to summarize the issue changes what it needs, while untrusted text is already in its context. Today, they note, an agent typically has uniform access to all tool calls, and they cite OpenAI's advice to enable only the connectors a task needs as least privilege enforced by hand. Existing work, they write, does not solve the problem of specifying the policy.
No app, no developer: least privilege for an AI agent has to come from the task, six security researchers argueBI-FIND-2025-0256
With its policy checks on, Microsoft's FIDES planner let 1 of 949 injection attacks succeed on gpt-4o in AgentDojo, against 163 for an undefended basic planner and 24 for FIDES without the checks. The checks do the blocking: under the same checks a basic planner let 9 through and a variable-passing planner 14, and all three reach zero once two injection tasks the authors set aside are dropped. One of those only changes what the agent says back to the user, and the authors write that "we only enforce policies upon tool calls, our planners do not stop these text-to-text attacks." In the other, the confidentiality policy allows an injected calendar event because no data leaves. So the evaluation's own claim is narrower than the introduction's "stops all prompt injection attacks": it blocks every attack that breaks its policies. That result also rests on labels the authors set, marking as untrusted every data field any benchmark injection targets.
In Microsoft's agent tests, label checks blocked policy-breaking tool calls and its FIDES planner finished more tasksBI-FIND-2025-0454
RTBAS, a defense from Carnegie Mellon authors for agents that call tools, blocks an injected call by checking where the call's data came from. Each message region carries an integrity label and a confidentiality label, a policy sets the label each tool requires, and a call that falls short waits for the user to confirm. The authors report they "prevent 100% of attacks that violate our security policy," but the paper assumes the labels and the policy are supplied, calls that an open problem and says it will likely burden agent developers. It also requires the tools themselves to comply with the policy: in the authors' example, a tool that returns sent messages as public could let a message sent to oneself "launder" private information. In the AgentDojo tests the authors assigned the integrity labels themselves, and in one workspace case user-written text labelled high-integrity held a possible injection and went untracked. Their own summary is that the guarantees are "only as good as the labels provided and the policies enforced." We read both as an access-policy job that needs an owner and a regular review. The paper's text never uses the words identity, credential or delegation, so its check looks at where data came from and leaves open who the agent acts for.
A Carnegie Mellon defense blocked every policy-breaking injection in tests, if its labels and tools hold upBI-FIND-2025-0647
Cited in 4 briefings
Every briefing that discusses this term, newest first.
- No app, no developer: least privilege for an AI agent has to come from the task, six security researchers argue
Paper · · arXiv · Mihai Christodorescu et al. · Dec 1, 2025
- In Microsoft's agent tests, label checks blocked policy-breaking tool calls and its FIDES planner finished more tasks
Paper · · arXiv · Manuel Costa et al. · May 29, 2025
- A Carnegie Mellon defense blocked every policy-breaking injection in tests, if its labels and tools hold up
Paper · · arXiv · Peter Yong Zhong et al. · Feb 13, 2025
- A Wisconsin agent defense stopped every benchmark injection by keeping all tool output from its planner
Paper · · arXiv · Fangzhou Wu et al. · Sep 27, 2024