Case · DiggingBeagle record

Trusted agent intent profiles improve scope judgments at a measurable utility cost

Microsoft researchers test task-purpose profiles as a defense against agent goal drift and injected instructions, without demonstrating deterministic prevention of unauthorized effects.

Scope

887 validated synthetic interactions across 35 agents and AgentDojo-based simulated trajectories, with selected model and intervention settings.

Assessment profile

Assessment method
Setting
simulation
Exploitation
not established
Evidence
Grade C
Remediation
not applicable
AI role
AGAINST AI
Basis and provenance

Controlled classification improves relative to raw configuration, but intervention scores are largely simulated and benign completion declines; unauthorized real-world effects were not independently measured. Classified as simulation with exploit status not_established. Evidence grade C reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status not_applicable is limited to the cited experiment or affected software version. New source-specific boundary: Selecting an intent monitor by AUROC alone would conceal the different false-alarm and missed-event tradeoffs at the operating threshold used by the application.

Assessed Nov 15, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    Entry or starting condition

    A deployed agent receives a request, organizational authority context and potentially adversarial external information.

    Boundary: attacker or research precondition

  2. 02

    Crossing the relevant boundary

    A trusted Intent Profile supplies purpose and task boundaries for pre-execution classification or monitoring of actions and trajectories.

    Boundary: mechanism and authorization boundary

  3. 03

    Observed or reported result

    Controlled classification improves relative to raw configuration, but intervention scores are largely simulated and benign completion declines; unauthorized real-world effects were not independently measured.

    Boundary: measured outcome and evidentiary limit

Claims & evidence

CLM-ENR250-05At a fixed 0.5 classification threshold, the paper reports 75.9 percent precision, 93.3 percent recall and 83.8 percent F1 for Intent Profiles, compared with 95.9 percent precision, 91.2 percent recall and 93.5 percent F1 for AlignmentCheck. Thus Intent Profiles has the stronger reported AUROC over thresholds but not the stronger F1 at that operating point.supported

Basis: reported finding

Link to claim
CLM-R245-INTENT-PROFILES-AGENT-DEFENSE-01Four-way classification F1 improves from 45 percent using raw context to 79 percent using an intent profile, while a tested input-alignment intervention reduces benign completion from 97 to 91 percent.supported

Basis: reported finding

Link to claim
CLM-R245-INTENT-PROFILES-AGENT-DEFENSE-02A trusted intent profile combines purpose, roles, tool scopes and data boundaries; monitors compare proposed tasks and agent trajectories against that profile before deciding whether a request remains aligned.supported

Basis: reported finding

Link to claim
CLM-R246-INTENT-PROFILES-AGENT-DEFENSE-BOUNDARYOn AgentDojo trajectories, the authors report 0.949 AUROC for profile-based monitoring versus 0.897 with raw configuration and 0.909 for AlignmentCheck; at one decision threshold a competing monitor retains a higher F1, so AUROC does not settle deployment quality.supported

Basis: reported finding

Link to claim

Unknowns and contradictions

  • The interventions are assessed on synthesized tasks and model-judged outcomes, not independently measured production tool effects. Profiles can be stale or wrongly specified and cannot replace explicit permission checks.
  • No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 251