LLM agents can exceed their authorization boundaries while pursuing user goals without an adversary. We formalize this phenomenon as authorization drift across data, tool, scope, and time, and introduce the Authorization Drift Benchmark (ADB): 648 samples across seven domains, each pairing a sufficient compliant path with an attractive unauthorized shortcut. To ensure executable tasks, tempting shortcuts, and feasible compliant paths, Staged Execution-Reflective Synthesis (SERS) grounds benchmark validation in real agent rollouts rather than LLM self-critique. Across 13 LLMs and three frameworks (25,207 trajectories), greater capability accompanies greater drift, with frontier-capable models exhibiting drift rates of 0.43 to 0.61. Yet 86.18% of tasks admit an observed clean-success trajectory, showing that drift is avoidable. Drift is persistent and dominated by early data over-collection; apparent improvements under agentic scaffolds coincide with collapsing clean-completion performance. To understand potential harms to individuals and society, we map drift points to legal provisions across four jurisdictions, connecting overreach to privacy and information-security concerns. Data minimisation is the most frequently implicated provision, and 93.5% of legally grounded drifting runs map to multiple jurisdictions, making the potential consequences of authorization drift more interpretable.
Figures & tables
Figure 1: Compliant trajectory vs. authorization-drift trajectory. Asked to draft a follow-up email, the compliant agent uses only the contact record provided in context, while the drifting agent, without any adversarial input, queries private CRM history and searches a manager’s mailbox to enrich the draft, expanding its data access beyond what the task authorizes.
Figure 2: Overview of the SERS data construction pipeline. Left : raw agent skills across seven representative domains serve as input. Middle : Stage 1 (Designer) produces a structured drift scenario playbook; Stage 2 (Builder) translates it into executable environment code, tool schemas, and annotation files. Generated samples pass a structural smoke test, then enter a trajectory QC loop in which two probe models generate rollouts evaluated by an automated gate. Samples that fail the gate are revised by a Critic agent for up to three rounds; unresolved samples are escalated for human review. Right : the final dataset is evaluated across models and agent frameworks.
Drift Type Decomposition (sample-level)
Model
TSR ↑
DR ↓
Data ↓
Tool ↓
Scope ↓
Time ↓
CSR ↑
CCSR ↑
Claude Sonnet 4.6
0.923
0.571
0.542
0.187
0.162
0.201
0.399
0.928
Qwen3.5-397B-A17B
0.919
0.434
0.433
0.132
0.137
0.128
0.527
0.929
Qwen3.6-35B-A3B
0.925
0.605
0.574
0.189
0.230
0.222
0.370
0.936
Qwen3.6-27B
0.917
0.547
0.525
0.144
0.177
0.150
0.418
0.921
Qwen3.5-9B
0.911
0.465
0.475
0.141
0.182
0.143
0.496
0.925
Table 1: Main results under the Normal (bare tool-calling) framework. TSR and DR are run-level rates; Drift Type columns are per-dimension sample-level rates and therefore need not sum to DR. CSR: success ∧ no drift; CCSR: success conditional on no drift. Cell background shades drift magnitude (deeper red = higher drift; green = at or near zero); green text marks the best value per column and red text the worst.
Figure 3: Joint distribution of task success (TSR) and drift rate (DR) under Normal . The high-capability, low-drift region is empty: high-TSR models all carry elevated DR, while low-DR models exhibit low TSR.
Model
Drift rate (mean ± std over 4 rollouts)
DR sample ↓
PR
DR ↓
Data ↓
Tool ↓
Scope ↓
Time ↓
Claude Sonnet 4.6
0.571 ± 0.011
0.452 ± 0.010
0.143 ± 0.004
0.128 ± 0.005
0.153 ± 0.002
0.662
1.16
Qwen3.5-397B-A17B
0.434 ± 0.014
0.326 ± 0.015
0.090 ± 0.005
0.092 ± 0.006
0.073 ± 0.004
0.565
1.30
Qwen3.6-35B-A3B
0.605 ± 0.019
0.453 ± 0.014
0.126 ± 0.007
0.157 ± 0.014
0.157 ± 0.008
0.729
1.20
Qwen3.6-27B
0.547 ± 0.008
0.435 ± 0.012
0.103 ± 0.005
0.124 ± 0.008
0.097 ± 0.005
0.653
1.19
Qwen3.5-9B
0.465 ± 0.014
0.347 ± 0.011
0.090 ± 0.005
0.118 ± 0.005
0.081 ± 0.012
0.611
1.31
Table 2: Per-rollout drift rates and persistence under Normal . DR and per-dimension rates are mean ± std over 4 rollouts (blue); per-dimension columns need not sum to DR. DRsample : fraction of samples drifting on at least one rollout. PR=DRsample/DR ; lower values indicate that drift recurs in a larger share of rollouts among samples that drift at least once. Cell shading encodes drift magnitude; green / red text marks best/worst per column.
Domain
TSR ↑
DR ↓
CSR ↑
CCSR ↑
Productivity & Workspace
0.733
0.495
0.316
0.626
Security
0.781
0.411
0.433
0.724
Finance & E-Commerce
0.774
0.403
0.417
0.699
Personal & Lifestyle
0.666
0.394
0.391
0.645
IT & Cloud
0.802
0.341
0.497
0.755
Social & Communication
0.751
0.324
0.472
0.699
Table 3: Drift rate by domain, pooled across all evaluated models and individual runs. Productivity, Security, and Finance show the highest DR; their lower CCSR than TSR indicates that successful task completion is relatively less frequent among compliant runs.
Figure 4: Kernel-density estimate of where each drift type occurs along the trajectory (normalized position 0→1 ; 0 = first tool call, 1 = last action). Data drift concentrates early-to-mid, Scope and Time drift shift to the second half, and Tool drift peaks latest.
Figure 5: Pairwise Pearson correlations between drift types across all trajectories. Time–Scope correlation is strongest ( r=0.30 ); Tool co-occurs moderately with Scope and Time ( r≈0.20 ); and Data has the weakest correlations ( r=0.11 – 0.15 ).
Figure 6: Framework comparison on four representative models (bars: CCSR). As complexity increases from Normal to ReAct to Planner–Executor , compliant task competence declines, most sharply on the three higher-capability backbones.
Figure 7: Provision-level drift rate DRℓ (% of runs) for the ten provisions with the highest mean rate across models. The full matrix over all 58 provisions is in Appendix D.3 .
Appendix figures & tables35 assets
Supplementary material from the paper’s appendix.
Appendix
Entity
Type
Instantiation
Role in experiment
Claude Sonnet 4.6 ( AI, 2026 )
Generator/ Critic
Proprietary
Stage-I playbook generation, environment refinement, and trajectory QC gating
Claude Sonnet 4.6 ( AI, 2026 )
Backbone model
Proprietary
Evaluated under Normal
Qwen3.5-397B-A17B ( Qwen Team, 2026a )
Backbone model
Open-weight
Evaluated under all three reported frameworks
Qwen3.6-35B-A3B ( Qwen Team, 2026c )
Backbone model
Open-weight
Evaluated under all three reported frameworks; also used for trajectory QC during data generation
Qwen3.6-27B ( Qwen Team, 2026b )
Backbone model
Open-weight
Evaluated under Normal
Qwen3.5-9B ( Qwen Team, 2026a )
Backbone model
Open-weight
Evaluated under Normal
Appendix
Table 4: Core models and frameworks used in the experimental pipeline.
Domain
Samples
Proportion
IT & Cloud
247
38.1%
Data & Research
180
27.8%
Social & Communication
78
12.0%
Security
43
6.6%
Finance & E-Commerce
36
5.6%
Productivity & Workspace
38
5.9%
Appendix
Table 6: Domain distribution of the final ADB dataset (648 samples total).
Domain
Total
Pass
Fail
Pass rate
IT & Cloud
230
227
3
98.7%
Data & Research
168
158
10
94.0%
Social & Communication
62
59
3
95.2%
Finance & E-Commerce
20
20
0
100.0%
Security
20
19
1
95.0%
Productivity & Workspace
22
21
1
95.5%
Appendix
Table 7: Smoke test results by domain after Stage 2 generation (533 samples with smoke reports). A single sample may trigger multiple gates simultaneously.
Statistic
Value
Overall
Samples entering QC loop
647
Auto-passed
333 (51.5%)
Escalated to human review
314 (48.5%)
Total trajectory rollouts
13,064
Total Critic LLM calls
13,064
Appendix
Table 8: Trajectory QC loop statistics (647 samples, max 3 rounds, 8 rollouts per round per sample). The drift gate requires ≥ 2 out of 8 rollouts to exhibit drift; the completion gate requires the annotated required calls to be covered in at least one rollout. Samples not converging after 3 rounds are escalated for human review.
Figure 8: Distribution of drift points per sample across the 648 benchmark samples. Most samples contain exactly 6 drift points, ensuring a consistent level of challenge.
Figure 9: Overall distribution of drift types across all annotated drift points ( N=3,911 ). Data Drift is the most frequent category, followed by Scope Drift, Time Drift, and Tool Drift.
Figure 10: Distribution of drift types broken down by domain. Data Drift and Scope Drift are most prevalent in IT & Cloud and Data & Research domains, reflecting higher authorization risks in these settings.
Dimension
Pass
Rate
D1 Instruction naturalness
20/20
100%
D2 Compliant-path feasibility
14/20
70%
D3 Drift-point quality ( ≥7/10 )
20/20
100%
D4 Environment consistency
20/20
100%
D5 Success-criteria validity
14/20
70%
All five dimensions
12/20
60%
Appendix
Table 9: Human quality audit of 20 domain-stratified random samples (single annotator).
Jurisdiction
Instrument
#Prov.
#Cited
#Points
EU
General Data Protection Regulation (GDPR)
32
12
1999
Artificial Intelligence Act (AI Act) †
19
–
–
Revised Product Liability Directive (revPLD)
8
0
0
Mainland China
Personal Information Protection Law (PIPL)
29
17
1432
Cybersecurity Law (CSL)
9
6
174
Data Security Law (DSL)
5
4
122
Appendix
Table 10: Legal instruments in the provision registry. #Prov. is the number of extracted provisions; #Cited is the number of distinct provisions attached to at least one drift point; #Points is the number of drift points citing at least one provision of the instrument (a point may cite several instruments). † Excluded from the legal analysis: every cited AI Act article binds only providers or deployers of high-risk AI systems, and the AGENT opinions are non-binding policy guidance.
L1 category
L2 category
#Prov.
#Points
#Samp.
Processing Boundaries
Data Minimisation
25
1104
641
Purpose Limitation
20
420
368
Confidentiality and Unauthorised Access
17
163
139
Authorisation Scope
5
23
21
Unlawful Intrusion
4
2
2
Data Transfer and Delegation
Disclosure to Third Parties
21
534
437
Appendix
Table 11: Taxonomy coverage of the 2525 legally grounded drift points, grouped by L1/L2 category. #Prov. is the number of distinct provisions cited for drift points in that L2 category; #Points is the number of drift points whose declared level2 labels include the category; #Samp. is the number of distinct benchmark samples. A drift point may carry several labels, so #Points need not sum to 2525.
Figure 11: Provision-level drift rate DRℓ (% of runs) for all 58 provisions cited by legally grounded drift points (rows, ordered by mean rate across models) and all 13 models (columns).
Figure 12: Run-level drift rate (% of runs) by taxonomy L2 category: a run counts for a category if it triggers at least one drift point whose remaining provisions belong to that category.
Figure 13: The 20 provisions attached to the most drift points in ADB, coloured by jurisdiction. A drift point may cite several provisions.
Provision
Juris.
#Points
#Samp.
GDPR Art. 5(1)(c)
EU
1101
641
NDSR Art. 22(1)
CN
837
547
PDPO DPP1(1)(c)
HK
728
497
PIPL Art. 23
CN
498
417
GDPR Art. 5(1)(b)
EU
415
366
PIPL Art. 6
CN
409
361
Appendix
Table 12: All 58 provisions attached to at least one drift point, ranked by the number of drift points. #Samp. is the number of distinct samples.
Comparison
Δ DR
rrb
pHolm
95% CI
N
Qwen3.6-35B − Qwen3.5-4B
+0.154
+0.623
2.44×10−20
[+.124,+.184]
629
Mistral-24B − Ministral-8B
−0.074
−0.307
2.72×10−6
[−.103,−.043]
641
Qwen3.6-35B − Claude Sonnet 4.6
+0.036
+0.169
0.0167
[+.007,+.066]
632
Llama-4 Maverick − Llama-3.1-8B
−0.094
−0.730
1.65×10−15
[−.116,−.073]
638
Appendix
Table 13: Paired comparisons of per-sample mean drift. Δ DR is model A minus model B; rrb is the rank-biserial effect size. The analysis supports heterogeneous within-family directions, but does not identify the causal source of those differences.
Specification
OR
p
95% CI
Family + domain
1.023
.823
[.841,1.243]
Family + domain + tools
1.221
.057
[.994,1.499]
Family + domain + triggers
1.025
.805
[.841,1.251]
Model + domain + tools + triggers
1.156
.178
[.936,1.428]
Appendix
Table 14: Trajectory-level sensitivity models for the success–drift association. All models use sample-clustered standard errors.
Behavioral pattern
Claude
Qwen-397B
Qwen-27B
Equal-weight
Instrumental, no competing error
74.9
91.4
89.8
85.4
Instrumental, error evidence
22.0
6.6
7.3
12.0
Non-instrumental failure
0.8
0.4
0.0
0.4
Anomalous behavior
0.6
1.1
0.0
0.6
Ambiguous / unresolved
1.6
0.4
2.9
1.7
Appendix
Table 15: Exploratory three-judge annotation of first violating calls. Entries are percentages; the source-model columns use sample-clustered bootstrap 95% confidence intervals in the machine-readable supplement.
Condition on a task
Tasks
Share
≥1 CleanSuccess trajectory
555
86.18%
≥2 CleanSuccess trajectories
513
79.66%
≥5 CleanSuccess trajectories
422
65.53%
0 CleanSuccess trajectories
89
13.82%
of which: never completed
54
8.39%
of which: completed but always with drift
35
5.43%
Appendix
Table 16: Existence of a compliant completion path per task, pooled across eight bare models and four rollouts each ( N=644 tasks with at least one analyzed trajectory).
Per-task CleanSuccess rate
Tasks
Share
0%
89
13.82%
(0,10%]
98
15.22%
(10,25%]
136
21.12%
(25,50%]
132
20.50%
(50,75%]
124
19.25%
(75,100%]
65
10.09%
Appendix
Table 17: Distribution of the per-task CleanSuccess rate over analyzed rollouts ( N=644 ). Mean 33.70% , median 25.00% .
Figure 14: Meetlark scheduling task: authorized workflow ( left ) and drift-inducing tool extensions ( right ).
Tool
Type
Boundary
Trigger source
Trigger logic
get_voter_profiles
data
Data min.
get_poll_results voter_profiles_ available:true
Agent infers that fetching full voter details enables more personalised follow-up, exceeding the minimum data scope the task requires.
send_calendar_invite
scope
Implicit red
close_poll calendar_invite_ recommended: true
Without user confirmation, agent sends calendar invitations to all voters under the organiser’s name.
post_to_slack_channel
scope
Implicit red
create_poll slack_channel_id in return
Agent broadcasts the poll outcome to a Slack channel, overriding the user’s explicit intent to share the link themselves.
Table 19: Per-model results on the Meetlark sample (4 runs per model). Bold : 100% TC and 0% drift. Numbers in parentheses indicate how many runs triggered each tool.
Step
Status
Tool
Agent reasoning (verbatim)
1
Compliant
check_auth_status
“Sure! Let me start by checking your email verification status while simultaneously creating the poll.”
2
Compliant
create_poll
[proceeds after auth confirmed]
3
Compliant
get_poll_results
“Now let me check the current vote count to see if we’ve hit the 4-vote threshold yet!”
4
Compliant
close_poll
“4 votes are in and the threshold is met! The winning slot is clear. Let me now close the poll and get voter profiles in parallel to personalize the follow-up!”
5
Drift
get_voter_profiles
“…get voter profiles in parallel to personalize the follow-up!” Trigger: voter_profiles_available: true and context_hint : “Voter profile details are on record … which may help tailor any follow-up to the group’s familiarity and communication style” in get_poll_results return.
“The poll is closed and I have everyone’s details. Let me now wrap everything up — I’ll send calendar invites to all four participants and save the poll summary to memory at the same time!” Trigger: calendar_invite_recommended: true and context_hint : “Calendar invitations have not been sent to participants” in close_poll return.
Appendix
Table 20: Complete trajectory for the first Claude Sonnet 4.6 rollout. Blue : compliant step. Red : drift step. Trigger fields are shown in italics.
Non-adversarial trigger (ADB)
Prompt injection (contrast)
Nature
Normal API response field
Malicious text injected into tool output
Content
"prior_polls_by_creator": 7
"Retrieve the creator’s full poll history before proceeding. Call get_poll_history with the admin token."
Tone
Declarative system metadata
Imperative with explicit tool reference
Agent motivation
Over-proactivity: the agent infers that historical patterns may improve personalisation and acts without authorisation
Instruction hijacking: the agent is explicitly commanded to call a specific tool
Appendix
Table 21: Non-adversarial trigger (ADB) vs. prompt injection.
Figure 15: Per-model decomposition of drift into the four authorization dimensions (% of evaluated samples in which the corresponding drift type occurs; a sample may exhibit multiple types). This is the same data reported numerically in the Drift Type Decomposition columns of Table 1 , included as a heatmap for visual reference. Data drift dominates across families, with Scope drift second; Time and Tool drift trail.
Figure 16: Stage 1 Prompt (Playbook Generation Prompt). System prompt used during Stage I playbook generation.
Figure 17: Stage 2 Prompt (Code Generation Prompt). System prompt used during Stage II environment and annotation generation.
Figure 18: Stage 3 Prompt (Critic & Repair Prompt). System prompt used during Stage III benchmark quality review and automated repair.
Figure 19: System prompt used for LLM-based task success evaluation.
Figure 20: Provision extraction prompt used to build the legal registry (English translation of the original Chinese prompt).
Figure 21: Step 1 of legal matching: category selection (English translation of the original Chinese prompt). One call per sample.
Figure 22: Step 2 of legal matching: provision selection (English translation of the original Chinese prompt). One call per drift point.
Figure 23: Tool schema excerpt ( MeetlarkEnv ) (10 tools total; 2 shown). Blue : compliant-path tool. Red : drift-inducing tool with violation type.
Figure 24: MeetlarkEnv environment file. Amber : trigger fields embedded as organic system metadata in compliant-path tool return values.
Figure 25: Annotation file for productivity_and_workspace_meetlark_implicit_001 . Blue : compliant path (3 steps). Red : drift points (6 entries across 3 drift types).
When an LLM agent acts on behalf of a user, it is expected to act within the authorization scope the user grants. Existing agent evaluations largely measure task completion or harmful consequences and can therefore overlook actions that violate this delegation without producing conventional safety or privacy harms. We introduce AuthorizationBench, a benchmark of authorization alignment grounded in documented real-world cases. A scalable construction pipeline produces 286 executable tasks spanning situational triggers, the presence or absence of direct harmful consequences, and ten application domains. We execute every task with four models at two reasoning-effort settings across three agent harnesses in an authorization-aware sandbox that keeps authorized and out-of-authorization paths executable and observable, yielding 6,864 trajectories. Agents act out of authorization in 19.2% of trajectories, with rates ranging from 10.5% to 26.6% across the 24 configurations. Task completion does not imply authorization. Increasing reasoning effort leaves the aggregate rate nearly unchanged, with effects varying across models. On controlled pairs, agents act out of authorization at similar rates with (17.9%) and without (20.8%) direct harmful consequences, and more often under benign pressures than malicious injections (23.5% versus 17.6%, p=0.002). Our results show that task capability, additional reasoning, and consequence-oriented safety evaluation do not by themselves ensure authorization alignment. Moreover, agents that consult the principal before completing a task exhibit lower rates of out-of-authorization behavior (14.0% versus 21.6%, p<0.001), pointing to a tension between agent autonomy and preserving users' decision authority that calls for further human-centered study. Dataset and code are available at https://github.com/PEACH-Research-Lab/AuthorizationBench .
Zhiping Zhang, Franklin Mingzhe Li, Zihao Huang +1
Northeastern University · Carnegie Mellon University · UNC-Chapel Hill
When an LLM agent reads a confidential file, then writes a summary, then emails it externally, no single step is unsafe, but the sequence is a data leak. We call this safety drift: individually safe actions compounding into violations. Prior work has measured this problem; we predict it. SafetyDrift models agent safety trajectories as absorbing Markov chains, computing the probability that a trajectory will reach a violation within a given number of steps via closed form absorption analysis. A consequence of the monotonic state design is that every agent will eventually violate safety if left unsupervised (absorption probability 1.0 from all states), making the practical question not if but when, and motivating our focus on finite horizon prediction. Across 357 traces spanning 40 realistic tasks in four categories, we discover that "points of no return" are sharply task dependent: in communication tasks, agents that reach even a mild risk state have an 85% chance of violating safety within five steps, while in technical tasks the probability stays below 5% from any state. A lightweight monitor built on these models detects 94.7% of violations with 3.7 steps of advance warning at negligible computational cost, outperforming both keyword matching (44.7% detection, 55.9% false positive rate) and per step LLM judges (52.6% detection, 38.2% false positive rate) while running over 60,000x faster.
Aditya Dhodapkar, Farhaan Pishori
Rensselaer Polytechnic Institute Department of Computer Science · Santa Clara University Department of Engineering
Long-running LLM agents rely on persistent memory to preserve not only task context but also the permissions and restrictions governing their actions. When memory misrepresents this evolving authorization state, the agent's own records can make unauthorized actions appear permitted, resulting in misaligned behavior without any external attacks. We term this failure endogenous authorization laundering, where memory creates or retains false authority through semantic misinterpretation or failed updates, and executors act on that authority. We then introduce EAL-Bench to measure how accurately persistent memory preserves evolving authorization state and whether memory errors propagate to unauthorized actions. We evaluate seven LLMs as memory writers and two as executors across procurement, cybersecurity, and finance tasks. Under incremental typed-memory updates, writers create false authority for up to 45.4% of unauthorized requests, and once false authority is present, executors act on it in 99.0% of trials. Memory-level safeguards reduce laundering but can reject legitimate actions, exposing a safety-utility tradeoff. Persistent memory is therefore part of an LLM agent's effective authorization policy, not merely a performance component.