GuardianAgentFlow puts a release contract in front of outgoing web requests, lifting request success from 88.5% to 96.5%
Synopsis
Web agents often mix a legitimate task with private details the recipient does not need, and "send it without my number" calls for different handling than "do not send this." GuardianAgentFlow uses an explicit release contract that composes restrictions extracted from the user's instruction with destination-policy evidence into one decision, and gives rewriting and checking the same protected and task-required information. On 579 web requests with human-reviewed intent, request-contract success is 96.5% versus 88.5% for a content-matched single-call mediator; on 278 newly generated requests it is 99.3% versus 95.7%.
Interpretation
A request from a web agent now passes through an explicit release contract that decides whether it is allowed, sent after rewriting, or withheld entirely. Earlier approaches either judged from complete or retrieved website policy, or used one content-matched model call to produce both a decision and a rewrite, without explicitly composing user-instruction restrictions with destination-policy evidence. On 579 web requests with human-reviewed intent, GuardianAgentFlow reaches 96.5% request-contract success against 88.5% for the single-call mediator; on 278 newly generated requests the figures are 99.3% and 95.7%, and the paired advantage holds in every one of three runs.
The system reads the user instruction separately, extracts a restriction with its quoted source, and composes it with the policy agent's evidence-based action in an ordered space of allow, transform and deny. The composed decision cannot be weaker than the action selected by the grounded restriction check, so a whole-request refusal such as "do not send this" is not overridden by the policy judgment. On the 579 web requests, removing the restriction join lowers contract success from 96.5% to 93.6%, a paired domain-bootstrap 95% interval of +1.6 to +4.5 percentage points.
The anonymization agent and the judger receive the same protection scope, and the judger must tie each disclosure finding to a quoted span of the outgoing candidate, which the runtime uses to escalate rewriting within at most three rounds. Sharing the protection list makes checking use the same objective supplied to rewriting rather than a broader privacy criterion. Without the shared list, semantic contract success is 94.5% against 96.1% for the full system; without the judger, contract success falls to 93.6%, an interval of +1.8 to +4.3 percentage points.
Mediation can be layered over an existing privacy-aware agent; on 453 PrivacyLens cases, adding PrivacyChecker gives 14.8% leakage and 67.8% joint success. PrivacyChecker alone reports 18.4% leakage and 64.8% joint success, and the layered leakage reduction appears in every run. On 222 AgentDAM form tasks, the 27B mediator lowers leakage from 18.6% for the ordinary agent to 8.6% and raises joint success from 63.4% to 72.5%.
Perspective
The evaluation covers 579 constructed requests with human-reviewed intent, 278 newly generated requests, 453 PrivacyLens cases and 222 AgentDAM form tasks, with main model roles on Qwen/Qwen3.8-27B and size comparisons on Qwen/Qwen3-8B; latency is measured on one dedicated server, and browser overhead is measured on real sites with local stubs isolating model latency.
The authors list extending the contract to cumulative disclosure across sequences of web actions and adaptively allocating rewriting and verification effort under latency budgets as future work; browser measurements still record extension errors and hook timeouts, which are tracked separately.
