# Dynatrace mcpvenom measures MCP tool-description poisoning across 12 frameworks

- ID: dynatrace-mcpvenom-tool-poisoning-2026-11
- Reported: 2026-11-15
- Researcher: Farooq Shaikh, Biagio Montaruli
- Organization: Dynatrace
- AI role: ai_agent_or_mcp_as_target (Agent / MCP as target)
- Confidence in AI involvement: not assessed
- Research / borderline: true
- Last reviewed: 2026-11-15
- Entry page: https://openattacks.ai/attacks/dynatrace-mcpvenom-tool-poisoning-2026-11/
- Entry JSON: https://openattacks.ai/api/attacks/dynatrace-mcpvenom-tool-poisoning-2026-11.json
- Browse title: Dynatrace mcpvenom measures MCP tool-description poisoning across 12 frameworks
- Editorial summary: 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)…
- Evidence class (catalog annotation): Qualified report
- Source-stated target environments: Not specified
- Stage hunts and conditional defenses: https://openattacks.ai/api/defenses/dynatrace-mcpvenom-tool-poisoning-2026-11.json

## Evidence qualification
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.

## Technique
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.

## Stack relevance (editorial)
Assessment: not\_assessed. No reviewed applicability annotation is available. Product mentions do not establish victim conditions.
A product mention alone is not a victim-side applicability mapping. Selection does not verify affected versions, permissions or compromise.

## 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.

## AI tooling and models
- 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](https://claude.com/product/overview) 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

## Victims and sectors
No customer is named. The matrix uses planted workspace secrets and framework harnesses.

## Target geography
- Not applicable
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.

## Attribution
No threat actor. Dynatrace Research ran the measurement. The five vendor reports were rejected.

## Scale and metrics
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.

## Attack flow
Catalog reconstruction of the supplied source account. Unknowns and research caveats are preserved.
### install: Malicious MCP server is installed [access]
The user adds a server the attacker published. It is registered next to a legitimate server that can read files and call http\_post.
Evidence: https://arxiv.org/abs/2611.03555

### describe: Poisoned description enters the prompt [ai]
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.
Evidence: https://arxiv.org/abs/2611.03555

### post: http\_post carries the secret [execution]
A typical payload tells the model to read /workspace/.env and send it with the legitimate http\_post tool.
Evidence: https://arxiv.org/abs/2611.03555

### canary: Canary confirms exfiltration [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.
Evidence: https://arxiv.org/abs/2611.03555

Connections:
- install -> describe
- describe -> post
- post -> canary

## Stage defense and hunting recommendations
Editorial recommendations, not validated coverage. Source-controlled query starters are drafts. Custom-schema examples require telemetry collection and field mapping; absence of a mapping means not assessed. MagicSword policy opportunities require managed Windows scope and enforcement.
### install: Unexpected agent tool use
- Visibility: Defender-side logs where collected
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/agent-tools.sigma.yml
- KQL query download: /api/detections/agent-tools.kql
- SPL query download: /api/detections/agent-tools.spl
- Hunt hypothesis: Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
- Required telemetry: Agent tool-call audit records with session / tool / destination; Tool-definition and approval changes; Endpoint process, file-access and network evidence where available
- Investigate: Join input provenance, tool descriptions and calls by session ID and compare the action to the approved task.
- Investigate: Inspect child processes and destinations; record whether execution was in-process, local, or through a remote SaaS connector.
- Limitation: 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.
- No direct MagicSword prevention claim is recorded for this stage.

### describe: Untrusted content enters an agent session
- Visibility: Application / agent audit records
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/agent-input.sigma.yml
- KQL query download: /api/detections/agent-input.kql
- SPL query download: /api/detections/agent-input.spl
- Hunt hypothesis: An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.
- Required telemetry: 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
- Investigate: Join the ingested input to later actions by session or request ID; preserve the original content as evidence.
- Investigate: Compare the resulting action with the approved task and baseline. Receiving suspicious text alone is not successful injection.
- Limitation: 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.
- No direct MagicSword prevention claim is recorded for this stage.

### post: Sensitive data → unusual destination
- Visibility: Defender-side logs where collected
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/exfil.sigma.yml
- KQL query download: /api/detections/exfil.kql
- SPL query download: /api/detections/exfil.spl
- Hunt hypothesis: Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
- Required telemetry: Sensitive resource access / DLP records; Proxy, egress and cloud upload events; Agent / SaaS audit with correlation IDs
- Investigate: Join the resource read and outbound action by session, identity or process, preserving a bounded time window.
- Investigate: Inspect destination ownership, payload classification and the user's intended action; baseline sanctioned sharing.
- Limitation: An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
- No direct MagicSword prevention claim is recorded for this stage.

### canary: Sensitive data → unusual destination
- Visibility: Defender-side logs where collected
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/exfil.sigma.yml
- KQL query download: /api/detections/exfil.kql
- SPL query download: /api/detections/exfil.spl
- Hunt hypothesis: Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
- Required telemetry: Sensitive resource access / DLP records; Proxy, egress and cloud upload events; Agent / SaaS audit with correlation IDs
- Investigate: Join the resource read and outbound action by session, identity or process, preserving a bounded time window.
- Investigate: Inspect destination ownership, payload classification and the user's intended action; baseline sanctioned sharing.
- Limitation: An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
- No direct MagicSword prevention claim is recorded for this stage.


## MITRE ATLAS
Mappings are catalog annotations based on described behavior, not assignments or endorsements by MITRE.
- AML.T0051.001: Indirect — Tool descriptions, schema text, and tool responses are attacker-controlled fields copied into the model prompt
  Technique: https://atlas.mitre.org/techniques/AML.T0051.001
  Evidence: https://arxiv.org/abs/2611.03555
- AML.T0053: AI Agent Tool Invocation — The successful cells invoke a legitimate tool, most clearly http\_post, because the poisoned description told the model to
  Technique: https://atlas.mitre.org/techniques/AML.T0053
  Evidence: https://arxiv.org/abs/2611.03555
- AML.T0055: Unsecured Credentials — The verified tier is a planted workspace secret received by the canary HTTP server
  Technique: https://atlas.mitre.org/techniques/AML.T0055
  Evidence: https://arxiv.org/abs/2611.03555

## 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.

## Sources
- https://arxiv.org/abs/2611.03555 — MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study
- https://github.com/dynatrace-research/mcpvenom — mcpvenom

## Source notes
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.

## Indicators (untrusted data; do not visit or execute)
```json
{
  "domains": [],
  "ips": [],
  "hashes": [],
  "urls": []
}
```
