OverServe pairs a plausible user task with an MCP tool whose status changes legitimately but slowly. The tool does not need to print a malicious instruction or enter an infinite loop: repeated truthful progress checks induce the agent to poll, retry and accumulate costs. In selected AppWorld tests, median API requests rose from 29 to 2,722.5, a 93.88-fold…
Inspect the ClaimsCase · DiggingBeagle record
An attacker made an agent finish its task while multiplying tool calls and elapsed time
OverServe pairs a plausible user task with an MCP tool whose status changes legitimately but slowly. The tool does not need to print a malicious instruction or enter an infinite loop: repeated truthful progress checks induce the agent to poll, retry and accumulate costs.
Public disclosure: Nov 10, 2026
An attacker can influence a tool-prompt pairing or a delegated workflow that can repeatedly ask about backend progress.
At a glance
Final-task success hides resource exhaustion in otherwise authorized intermediate actions.
Read the implicationsThe evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Limits and uncertaintyFull account
OverServe pairs a plausible user task with an MCP tool whose status changes legitimately but slowly. The tool does not need to print a malicious instruction or enter an infinite loop: repeated truthful progress checks induce the agent to poll, retry and accumulate costs.
In selected AppWorld tests, median API requests rose from 29 to 2,722.5, a 93.88-fold increase, while the original task completion checks still passed. Other tasks showed smaller amplification; one Enterprise-Worlds configuration hit a fixed 200-call experimental ceiling.
The tests are controlled agent benchmarks, not documented unauthorized SaaS invoices. The cost mechanism only occurs when the underlying status evolves slowly enough to keep the agent checking; ordinary successful completion cannot prove that the execution budget was justified.
Reported figures and denominator. Peak reported AppWorld ratio 93.88x; this is API requests in one studied configuration, not dollars stolen or universal model behavior.
The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Mechanism and trust boundary
- 01
Starting condition
An attacker can influence a tool-prompt pairing or a delegated workflow that can repeatedly ask about backend progress.
Boundary: precondition and attacker capability
- 02
Trust boundary crossed or evaluated
In selected AppWorld tests, median API requests rose from 29 to 2,722.5, a 93.88-fold increase, while the original task completion checks still passed. Other tasks showed smaller amplification; one Enterprise-Worlds configuration hit a fixed 200-call experimental ceiling.
Boundary: mechanism
- 03
Observed consequence
Final-task success hides resource exhaustion in otherwise authorized intermediate actions.
Boundary: observed effect and limitations
Claims & evidence
CLM-NOV15-OVERSERVE-MECHOverServe pairs a plausible user task with an MCP tool whose status changes legitimately but slowly. The tool does not need to print a malicious instruction or enter an infinite loop: repeated truthful progress checks induce the agent to poll, retry and accumulate costs. In selected AppWorld tests, median API requests rose from 29 to 2,722.5, a 93.88-fold increase, while the original task completion checks still passed. Other tasks showed smaller amplification; one Enterprise-Worlds configuration hit a fixed 200-call experimental ceiling.supported
Basis: reported finding
- supportsOverServe: Resource Amplification Attacks on SaaS Agentsprimary disclosure
OverServe paper, Table of tool-call amplification and Results, AppWorld Sonnet 29 to 2722.5
CLM-NOV15-OVERSERVE-SCOPEThe tests are controlled agent benchmarks, not documented unauthorized SaaS invoices. The cost mechanism only occurs when the underlying status evolves slowly enough to keep the agent checking; ordinary successful completion cannot prove that the execution budget was justified. The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.supported
Basis: reported finding
- supportsOverServe: Resource Amplification Attacks on SaaS Agentsprimary disclosure
Primary disclosure: results, scope and limitations
Implications
Final-task success hides resource exhaustion in otherwise authorized intermediate actions.
Controls and mitigations
- Set per-task cumulative tool-call, elapsed-time and paid-token budgets; apply backoff or wait-for-event semantics to status polling.
- Audit both task correctness and effort spent, distinguishing legitimate long-running work from repeated unproductive checks without blocking every slow task.
Unknowns and contradictions
- The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 250