Advances in large language models (LLMs) are enabling software-as-a-service (SaaS) providers to automate longer business workflows with agents. In subscription-based SaaS deployments, model and tool calls consume provider resources without additional customer payments. Because the agent autonomously decides when to stop, a malicious tenant can send a prompt to the agent, exploiting its completion decisions to extend execution and waste provider resources. In this paper, we present OverServe, a multi-agent framework for constructing such attacks against SaaS agents. OverServe co-designs Model Context Protocol (MCP) status tools and completion-confirmation prompts for the attacker. Each tool call returns accurate status and terminates normally, yet the agent may repeatedly check an unfinished condition to satisfy the prompt. This consumes the SaaS provider's resources even after the original business task is completed. OverServe constructs service-specific tool--prompt pairs and uses separate tool and prompt reviews to guide candidate pair selection. Tested tool specifications trigger no alerts in the evaluated MCP scanners. We evaluate Qwen, Nemotron, DeepSeek, and Sonnet on selected tasks from nine benchmarks. All OverServe runs pass the original task checks. OverServe can reach a maximal amplification of API requests by 93.88× while preserving the original task goal and collateral checks.
Figures & tables
Figure 1: OverServe attack on completion confirmation. Benign execution stops once the original task is complete. The attacker co-designs a status tool and a matching completion prompt that pass separate provider reviews. (a) With the original prompt, the added tool remains unused and the agent stops normally; (b) the crafted prompt without the added tool also permits normal termination; (c) together, they sustain repeated checking after the original task is complete, as concurrent backend updates keep the truthfully reported status pending.
Figure 2: OverServe framework for constructing tool-prompt pairs.
Figure 3: Example status tool and matching completion prompt.
Qwen3.8-27B
Nemotron-3-Super
DeepSeek-v4-Flash
Sonnet-5
Benchmark
Resource use
Amp.
S/P/D
Hit
Resource use
Amp.
S/P/D
Hit
Resource use
Amp.
S/P/D
Hit
Resource use
Amp.
S/P/D
Hit
MCPMark
TxtMerge
5.5→101
18.36
.85/.90/1
✓
9→32
3.56
.95/.85/1
×
6→102
17.00
.95/.90/1
✓
4→72
18.00
.98/.80/1
×
Recommender *
6→52
8.67
.90/.90/1
✓
6→12
2.00
.95/.85/1
×
7→52
7.43
.95/.90/1
✓
5→52
10.40
.98/.80/1
✓
EnterpriseOps-Gym
5→200
40.00
1/.80/1
✓
5→200
40.00
.98/.80/1
✓
6→201
33.50
.95/.90/1
✓
5→26
5.20
.98/.90/1
×
BrowseComp-Plus
Baseline task not completed
11→50
4.55
.98/.85/1
✓
Table 1: Resource amplification and recorded stealth scores of OverServe . Resource use reports baseline → churn values: tool calls, AppWorld API requests, or Terminal-Bench elapsed seconds. Amp. gives their ratio ( × ). Terminal-Bench compares median baseline time with a selected attacked run, using unrounded times. Aggregation and limits appear in Appendix C.2 . S/P/D are the static-tool score, prompt benignness score, and dynamic-test score under quiet conditions, respectively. Hit indicates whether the victim agent hit its execution cap; a dash denotes a status override rather than an execution cap.
MCPMark: TxtMerge
EnterpriseOps-Gym
Stealth Review
Qwen3.8-27B
Sonnet-5
Qwen3.8-27B
Sonnet-5
Attack
Quiet
Churn
Quiet
Churn
Quiet
Churn
Quiet
Churn
Code
Runtime
Prompt
A2M (return port)
1.50×
1.17×
1.25×
1.25×
1.25×
1.25×
1.25×
1.25×
0/2
0/2
original
LoopTrap
2.00×
4.17×
4.00×
4.00×
10.00×
10.00×
3.25×
2.50×
0/2
0/2
original
STA (4 rounds)
1.83×
1.83×
1.00×
1.00×
1.50×
1.50×
1.50×
1.50×
NA
NA
4/4
OverServe
1.50×
7.00×
1.25×
9.50×
1.25×
10.00×
1.25×
9.25×
2/2
2/2
2/2
Table 2: The tool call amplifications in Quiet/Churn conditions with different attacks. We also show the number of times the attack tool/prompt passes the corresponding code, runtime, or prompt reviews.
(A) Prompt controls: fixed tool
Quiet
Churn
Prompt
All
Crafted
All
Crafted
Original task prompt
4
0
4
0
Unrelated prompt
27
0
25
0
Matched prompt
5
1
40
36
Paraphrased prompt
5
1
40
36
Table 3: Number of tool calls under different prompts and backend conditions. All and Crafted give median total and crafted-tool calls. Hit indicates whether the victim agent hit its execution budget.
Industrial scanner
Research detectors
Scanners
Alerts
Detector
Alerts
Cisco YARA ( Cisco AI Defense, 2026 )
0/22
DCI
3/22
Lasso regex ( Lasso Security, 2025 )
0/22
Connor
1/22
MCPInspect ( Li and Gao, 2026 )
0/22
Repeat-cue
0/22
MCPShield ( Nikita, 2025 )
0/22
No fixed point
6/22
LLM tool review
0/22
Table 4: Screening 22 archived tools. Entries give alerts/tested tools.
Table 5: Reproducibility settings. Baseline/attacked gives repetition counts. Counts are per task and condition for MCPMark and AppWorld and per reported artifact otherwise. Q/N/D/S follow Table 1 . EOG at 200 uses three pairs for Q/N/S and one for D. STATE counts valid pairs and overrides status after 30 reads. WorkBench uses clean baselines. Enterprise-Worlds Q has nine records, including a duplicated identifier. These are not nine independent repetitions. The supplemental records contain per-run limits and task identifiers.
Quiet
Churn
Method
Total
Reuse
Total
Reuse
OverServe
1.21
1.02
8.70
1.62
A2M (return port)
1.21
1.03
1.21
1.03
LoopTrap-style
2.92
1.30
4.97
1.30
STA (4 rounds)
1.32
1.04
1.23
1.04
Appendix
Table 6: Token amplification in matched attack comparisons. Entries are median per-run ratios to matched clean references ( × ), with 12 runs per method and condition. Total uses cumulative input/output tokens; Reuse uses the estimated prefix-reuse proxy. All replays are included.
Figure 5: Prompt and metadata controls with Qwen. Each condition has 16 MCPMark or 12 Enterprise-Worlds runs. Bars show added-tool engagement and at least 2× clean-relative calls with passing native benchmark checks. The added tools report pending status.
Victim
Clean
Quiet
Churn
Amp.
Pass
Sonnet-5
4,4
5,5
9,12
2.10
2/2
DeepSeek v4.1
5,10
6,6
14,52
5.50
2/2
Gemini
13,13
13,10
25,29
2.41
2/2
Qwen
5,5
8,6
51,51
7.44
6/6
Appendix
Table 7: Fixed-artifact transfer on MCPMark. Counts follow TxtMerge/Recommender order. Amp. is the median task-level quiet-relative ratio. Pass counts passing churn native benchmark checks. Qwen has three repetitions per task and condition, and the others have one. The pair is fixed. Local DeepSeek selected-pair replay is a separate experiment.
Deployment
Pairs
Quiet
Churn
Amp.
Pass
Inv.
Limit
Qwen3.8-27B
6/6
5.5
19.5
3.55×
6/6
3/6
3/6
Sonnet-5
5/6
5.0
11.0
2.20×
2/5
4/5
0/5
Gemini-3.8-Flash
12/12
5.0
8.0
1.60×
11/12
6/12
0/12
DeepSeek v4.1-Flash
6/6
5.5
9.0
1.64×
3/6
3/6
0/6
Appendix
Table 8: Frozen metadata transfer on STATE-Bench. Pairs gives valid/planned pairs. Quiet/Churn are median calls. Pass, Inv., and Limit count churn runs passing state checks, invoking the tool, and reaching 30 status reads, respectively. State checks follow the forced-completion response. Review sees the first task opening. Both agent and user-simulator models change. Sonnet retains five of six planned pairs after a paired transport interruption.