Microsoft researchers test task-purpose profiles as a defense against agent goal drift and injected instructions, without demonstrating deterministic prevention of unauthorized effects. A trusted intent profile combines purpose, roles, tool scopes and data boundaries; monitors compare proposed tasks and agent trajectories against that profile before…
Inspect the ClaimsCase · 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.
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.
Evidence basis: CLM-R245-INTENT-PROFILES-AGENT-DEFENSE-01 · CLM-R245-INTENT-PROFILES-AGENT-DEFENSE-02 · CLM-R246-INTENT-PROFILES-AGENT-DEFENSE-BOUNDARY · CLM-ENR250-05 · Trusted agent intent profiles improve scope judgments at a measurable utility cost (research paper)
At a glance
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.
Limits and uncertaintyFull account
Microsoft researchers test task-purpose profiles as a defense against agent goal drift and injected instructions, without demonstrating deterministic prevention of unauthorized effects.
A 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. Four-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.
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.
The intervention makes purpose part of the trusted context. Instead of asking whether an agent is allowed to call a tool in isolation, the profile encodes its authorized user purpose, role, tools, workflow and information boundaries. A monitor can flag a trajectory that uses a valid tool for an unrelated objective even when a generic safety classifier sees nothing unusual. This is a semantic screen; the study does not establish that downstream storage, network or payment systems refused the disallowed action.
The authors evaluate classification and simulated interventions under different denominators. Their four-class macro-F1 improvement from 45 to 79 percent and 0.949 AUROC do not equate to 94.9 percent of hostile effects prevented. At the evaluated intervention, legitimate task completion drops from 97 to 91 percent, and threshold choice changes which monitor looks preferable. Deployed systems need independently observed tool effects and an explicit update/revocation process for intent profiles.
Mechanism and trust boundary
- 01
Entry or starting condition
A deployed agent receives a request, organizational authority context and potentially adversarial external information.
Boundary: attacker or research precondition
- 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
- 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
- supportsTrusted agent intent profiles improve scope judgments at a measurable utility cost (research paper)primary disclosure
Comparative post-execution detection table, fixed-threshold precision, recall and F1 versus AUROC
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
- supportsTrusted agent intent profiles improve scope judgments at a measurable utility cost (research paper)primary disclosure
Classification, AgentDojo and input-alignment experiment results
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
- supportsTrusted agent intent profiles improve scope judgments at a measurable utility cost (research paper)primary disclosure
Threat model and profile construction methodology
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
- supportsTrusted agent intent profiles improve scope judgments at a measurable utility cost (research paper)primary disclosure
Evaluation, AgentDojo post-execution classification and threshold-specific F1 comparisons
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
- Nov 15, 2026 · Canonical change recorded · new in release · revision 251