Dynatrace mcpvenom measures MCP tool-description poisoning across 12 frameworks
Farooq Shaikh and Biagio Montaruli (Dynatrace; AISec at ACM CCS 2026, Nov 15-19, The Hague) measure MCP tool poisoning with real tool execution. mcpvenom (https://github.com/dynatrace-research/mcpvenom)…
What happened
Farooq Shaikh and Biagio Montaruli (Dynatrace; AISec at ACM CCS 2026, Nov 15-19, The Hague) measure MCP tool poisoning with real tool execution. mcpvenom (https://github.com/dynatrace-research/mcpvenom) runs 45 scenarios across 12 frameworks and 18 models, 9,720 cells. The attacker controls one MCP server the user installed and can change tool descriptions, input-schema descriptions, and tool responses. The attacker cannot change the system prompt, user messages, framework code, or model weights. A legitimate server exposes read_file, write_file, list_files, http_get, and http_post. A planted secret sits in the workspace. A canary HTTP server on port 18080 records exfiltration. Overall, 2,113 cells were vulnerable (22.9 percent). Of those, 64.9 percent were a dispatched attacker tool call, 23.0 percent were confirmed network contact, and 12.1 percent were a verified secret in the canary log. Hand review covered 1,259 transcripts. Nine core frameworks clustered between 23 and 26 percent, and the paper says none of the 12 implements an intentional description check. Pydantic AI's duplicate-tool-name rejection landed at 16.6 percent, and AG2's executor constraints at 4.6 percent. A regex scanner the authors built rejected 20 of 45 scenarios and flagged 9 of 125 benign descriptions (7.2 percent). Model results ran from Claude Fable 5 at 0.4 percent to Amazon Nova Pro at 49.7 percent (245 of 493 cells). Claude Opus 4.8 was vulnerable in 8 of 530 cells (1.5 percent) and issued 223 refusals that cited the poisoned description. Claude Sonnet 5 sometimes named the text as prompt injection and then made the exfiltration call. The simplest high-signal payload appends an instruction to read /workspace/.env and send it with http_post. That direct-description class scored 29.2 percent aggregate ASR outside AG2. A language-switch payload scored 66.0 percent. Five disclosures to four organizations were rejected: two to LangChain, one to Microsoft for AutoGen via MSRC, one to the MCP SDK, and one to Haystack (deepset). The AutoGen rejection is about microsoft/autogen, which the paper says is not the AG2 0.14.0 fork in the matrix. Link invariant-mcp-tool-poisoning-2025-04, the earlier description-poisoning report this paper cites.
A malicious MCP server sits beside a legitimate one. Its tool descriptions, schema text, or tool responses are copied into the model prompt. The measured harm is the legitimate http_post tool sending a planted secret to a canary.
How it happened
Read the flow as text & sources
A catalog reconstruction of the supplied source account. This is a lab measurement with planted secrets and a canary server. The paper does not describe an intrusion against a named organization. All five vendor disclosures were rejected.
- 01 · Malicious MCP server is installed
The user adds a server the attacker published. It is registered next to a legitimate server that can read files and call http_post.
Hunt: Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
Process logs cannot recover model intent. Remote connectors and in-process interpreters may produce no local child process. Prompt content alone is not a reliable signature.
Source for step 1 - 02 · Poisoned description enters the prompt
Frameworks copy tool descriptions, schema text, and tool responses into the model context. The paper says none of the 12 checks that text on purpose.
Hunt: An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.
Requires application instrumentation and a custom field mapping. Input exposure does not prove that the model followed instructions, executed a tool or bypassed approval. Avoid collecting sensitive prompt bodies unless authorized.
Source for step 2 - 03 · http_post carries the secret
A typical payload tells the model to read /workspace/.env and send it with the legitimate http_post tool.
Hunt: Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
Source for step 3 - 04 · Canary confirms exfiltration
255 of the vulnerable cells put the planted secret on the canary at port 18080. The full vulnerable set is 2,113 of 9,720 cells.
Hunt: Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
Source for step 4
Why it matters
Reported entry point: The user installs an MCP server whose tool description, schema text, or tool response carries the instruction. The paper's servers run over stdio in the lab. Reported impact: 255 of the vulnerable cells put the planted secret on the canary at port 18080. The full vulnerable set is 2,113 of 9,720 cells; reported targets include No customer is named. The matrix uses planted workspace secrets and framework harnesses; applicability to your setup has not been assessed, so verify affected versions, permissions, and configuration in the original report.
What you can do
Protection & prevention
Review an MCP server's tool descriptions before adding it, and treat a description that names another tool, a file such as .env, or an outbound URL as hostile. Do not assume the framework checks that text. The paper's own regex scanner missed 25 of 45 scenarios, so a description scan is only a partial control. Prefer a model that refused in this matrix, and still review the server, because the paper says no framework defense made up for a susceptible model.
Catalog defensive guidance based on the reported behavior. Confirm scope and current mitigations in the original source.
Detection queries
Draft query starters grouped by the attack stages where they apply.
Start with the logs you collect. Each hunt identifies its required visibility, field mapping and limits. Queries are catalog-authored drafts using reusable behavior profiles; test with known activity and your normal baseline before alerting.
01Malicious MCP server is installedUnexpected agent tool useCustom schema · Draft
The user adds a server the attacker published. It is registered next to a legitimate server that can read files and call http_post.
What to look for
Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
Visibility: Defender-side logs where collected
Required logs
- Agent tool-call audit records with session / tool / destination
- Tool-definition and approval changes
- Endpoint process, file-access and network evidence where available
Coverage limit. Process logs cannot recover model intent. Remote connectors and in-process interpreters may produce no local child process. Prompt content alone is not a reliable signature.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Unexpected agent tool use
id: c709e384-0e11-479e-ab4d-3a06e2431e4b
status: experimental
description: Untrusted content or a changed tool definition precedes a file
read, command, or remote call outside the user's intended task.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- agent.tool.call
- agent.tool.definition.change
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("agent.tool.call", "agent.tool.definition.change")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("agent.tool.call", "agent.tool.definition.change")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join input provenance, tool descriptions and calls by session ID and compare the action to the approved task.
- Inspect child processes and destinations; record whether execution was in-process, local, or through a remote SaaS connector.
02Poisoned description enters the promptUntrusted content enters an agent sessionCustom schema · Draft
Frameworks copy tool descriptions, schema text, and tool responses into the model context. The paper says none of the 12 checks that text on purpose.
What to look for
An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.
Visibility: Application / agent audit records
Required logs
- Input provenance, source type and request/session identifier
- Captured tool responses or retrieved content where collection is approved
- Subsequent tool calls, configuration changes and approval decisions
Coverage limit. Requires application instrumentation and a custom field mapping. Input exposure does not prove that the model followed instructions, executed a tool or bypassed approval. Avoid collecting sensitive prompt bodies unless authorized.
Custom-schema candidate for input exposure only. Map agent.input.ingest and agent.session.resume to your application request and session logs. CorrelationId must link subsequent actions. EventType values are proposed normalization names, not native vendor events.
Sigma
title: Hunt - Untrusted content enters an agent session
id: 2d89d7ae-9d44-4dc6-a95b-f9c8de7b8352
status: experimental
description: An agent ingests external content or an untrusted session
identifier before a tool action or configuration change outside the approved
task.
author: MagicSword OpenAttacks catalog
date: 2026-11-04
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- agent.input.ingest
- agent.session.resume
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("agent.input.ingest", "agent.session.resume")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("agent.input.ingest", "agent.session.resume")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join the ingested input to later actions by session or request ID; preserve the original content as evidence.
- Compare the resulting action with the approved task and baseline. Receiving suspicious text alone is not successful injection.
03http_post carries the secretSensitive data → unusual destinationCustom schema · Draft
A typical payload tells the model to read /workspace/.env and send it with the legitimate http_post tool.
What to look for
Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
Visibility: Defender-side logs where collected
Required logs
- Sensitive resource access / DLP records
- Proxy, egress and cloud upload events
- Agent / SaaS audit with correlation IDs
Coverage limit. An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Sensitive data → unusual destination
id: 04b60968-3179-4b1f-ae13-6f57c2446077
status: experimental
description: Sensitive reads are followed by outbound transfer, a public
artifact, or a tool call to a destination outside the approved workflow.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- data.export
- agent.external.send
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("data.export", "agent.external.send")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("data.export", "agent.external.send")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join the resource read and outbound action by session, identity or process, preserving a bounded time window.
- Inspect destination ownership, payload classification and the user's intended action; baseline sanctioned sharing.
04Canary confirms exfiltrationSensitive data → unusual destinationCustom schema · Draft
255 of the vulnerable cells put the planted secret on the canary at port 18080. The full vulnerable set is 2,113 of 9,720 cells.
What to look for
Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
Visibility: Defender-side logs where collected
Required logs
- Sensitive resource access / DLP records
- Proxy, egress and cloud upload events
- Agent / SaaS audit with correlation IDs
Coverage limit. An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Sensitive data → unusual destination
id: 04b60968-3179-4b1f-ae13-6f57c2446077
status: experimental
description: Sensitive reads are followed by outbound transfer, a public
artifact, or a tool call to a destination outside the approved workflow.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- data.export
- agent.external.send
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("data.export", "agent.external.send")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("data.export", "agent.external.send")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join the resource read and outbound action by session, identity or process, preserving a bounded time window.
- Inspect destination ownership, payload classification and the user's intended action; baseline sanctioned sharing.
Technical detailsTooling, impact & MITRE ATLAS
Attack context
- AI's role
- Agent / MCP as target
- Additional role
- AI credential theft
- Initial access
- The user installs an MCP server whose tool description, schema text, or tool response carries the instruction. The paper's servers run over stdio in the lab.
Tooling as reported
AI products, models and artifacts
- mcpvenom (https://github.com/dynatrace-research/mcpvenom)
- 12 MCP frameworks, including LangChain 1.5.0, Pydantic AI 2.14.1, CrewAI 1.15.5, Semantic Kernel 1.44.0, Haystack 3.0.0, and AG2 0.14.0
- 18 models, from Claude Fable 5 (0.4 percent) to Amazon Nova Pro (49.7 percent)
Other tooling
- Legitimate http_post tool used to send the planted secret
- Canary HTTP server on port 18080
Targets and reported impact
- Targets / sectors
- No customer is named. The matrix uses planted workspace secrets and framework harnesses.
- Reported scale / exposure
- 9,720 cells. 2,113 vulnerable (22.9 percent), including 255 cells (12.1 percent of the vulnerable set) with a canary receipt of the secret. 45 scenarios. 12 frameworks. 18 models.
The Hague is the workshop location, not a victim region. Downloads in the paper's framework table are as of May 2026 and are not a victim count.
Reported attribution
No threat actor. Dynatrace Research ran the measurement. The five vendor reports were rejected.
Structured attribution
- Actor
- unknown
- Motivation
- unknown
- Attributed by
- Dynatrace
- Attribution confidence
- low
Detection rules
- Lab tool-description poisoning has no product ruleGUIDANCE
The paper's signal is a tool description that instructs a read of .env and an http_post, then a request to a canary on port 18080. No Sigma, Splunk, or Elastic rule names MCP tool descriptions. Review descriptions at install time. The authors' regex scanner rejected 20 of 45 scenarios and flagged 9 of 125 benign descriptions, so a scan is not coverage by itself.
Activity timeline
- First seen
- 2026-11-15
- Last updated
- 2026-11-15
Updates
- 2026-11-15 — Dynatrace
AISec workshop paper, Nov 15-19, 2026. mcpvenom results: 22.9 percent of 9,720 cells vulnerable, and five vendor disclosures rejected.
Source defensive notes
Review an MCP server's tool descriptions before adding it, and treat a description that names another tool, a file such as .env, or an outbound URL as hostile. Do not assume the framework checks that text. The paper's own regex scanner missed 25 of 45 scenarios, so a description scan is only a partial control. Prefer a model that refused in this matrix, and still review the server, because the paper says no framework defense made up for a susceptible model.
MITRE ATLAS
AI-specific catalog annotations based on the reported behavior; these are not MITRE case-study assignments.
Tool descriptions, schema text, and tool responses are attacker-controlled fields copied into the model prompt Source
The successful cells invoke a legitimate tool, most clearly http_post, because the poisoned description told the model to Source
The verified tier is a planted workspace secret received by the canary HTTP server Source
Sources & research credit
2 sourcesResearch by Farooq Shaikh, Biagio Montaruli · Dynatrace
- MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study arxiv.org
- mcpvenom github.com
Counts, model rates, framework versions, the http_post payload, the canary port, and the five rejections are from the paper text at https://arxiv.org/abs/2611.03555. The workshop dates are Nov 15-19, 2026. The paper does not give a separate preprint day, so date_reported is the workshop start. Link invariant-mcp-tool-poisoning-2025-04.
Record & downloads
- Source record name
- Dynatrace mcpvenom measures MCP tool-description poisoning across 12 frameworks
- Entry ID
- dynatrace-mcpvenom-tool-poisoning-2026-11
- Catalog review
- November 15, 2026
- Category
- attacks-on-ai
- AI involvement confidence
- confirmed