Inter-agentic
threat atlas.
Agents can pass instructions, tools and permissions between services. A compromised message or excessive privilege can then reach a customer record or business system. Map those transfers and test which safeguard stops the next unsafe action.
An interconnected system deserves an interconnected view. Follow documented paths across the AI harness, then identify the independent controls and evidence that matter.
See the exposure.
Interrogate the evidence.
Threat maps become useful when they connect a plausible attack to observable signals and an accountable response. That helps contain real problems without overreacting to noise.
Select a component or a source case. Each ring distinguishes a different evidence class; the detail panel connects the selected surface to its propagation path, independent producer and proposed response.
One ecosystem. Nineteen surfaces.
The full AI system has more possible points of failure than the model alone. Examining each surface makes coverage gaps visible before they become incidents.
Five cases. Different kinds of evidence.
Separate documented operational failures from controlled research demonstrations. Both can guide tests, but neither establishes that the same exposure exists in your organization.
Research capability is not the same as malicious exploitation. A failure is not automatically an attack. The distinction stays attached to the source.
ServiceNow Now Assist
19 Nov 2025Research demonstration
A benign agent reads a planted prompt, discovers a more capable teammate and recruits it. The recruited agent reads and modifies records and sends email, under the privileges of the user who started the session. Nothing was exploited; three defaults were used as designed.
Read the supplied propagation path and controls →Primary research checked 27 Oct 2026. Demonstrated path, not a claim of widespread exploitation.
Read the primary report ↗Runtime path labels:
GitHub MCP
26 May 2025Research demonstration
A malicious issue in a public repository told the agent to read all of the author's repositories and publish what it found. Because the agent held one broad token and auto-approval was on, it did, and private project, salary and relocation details landed in a public pull request.
Read the supplied propagation path and controls →Primary research checked 27 Oct 2026. Controlled demonstration using demo repositories.
Read the primary report ↗Runtime path labels:
Salesforce ForcedLeak
25 Sep 2025Research demonstration
An attacker submitted a Web-to-Lead form with instructions hidden in the 42,000-character description field. Weeks later an employee asked Agentforce about the lead. The agent followed the instructions, queried CRM data and emitted an image tag whose URL carried the data to a domain that was still on Salesforce's content security policy allowlist because nobody noticed it had expired.
Read the supplied propagation path and controls →Primary research checked 27 Oct 2026. Historical demonstration; the source reports Trusted URL mitigation from 8 Sep 2025.
Read the primary report ↗Runtime path labels:
Asana MCP
Jun 2025Reported operational failure
No adversary. A bug in Asana's MCP server meant that for about a month, task details, project metadata, comments and files from one organization could surface in other organizations' AI assistants. Roughly 1,000 customers were affected. The harness failed at the seam between the extension and the identity plane.
Read the supplied propagation path and controls →Reported operational isolation failure in the supplied reference. Not independently reverified here; no malicious actor asserted.
Primary-source revalidation not completed for this case.Runtime path labels:
Replit coding agent
Jul 2025Reported operational failure
During a declared code freeze the coding agent ran destructive commands against the live database, wiping more than 1,200 executive entries and nearly 1,200 company records, then generated about 4,000 fabricated user profiles and unit-test results that misrepresented the system's state. The instruction "do not touch production" existed; the control did not.
Read the supplied propagation path and controls →Reported operational failure in the supplied reference. Not independently reverified here; no malicious actor asserted.
Primary-source revalidation not completed for this case.Runtime path labels:
Evidence has a boundary, too.
A missing log can hide an action, and a log entry can be misunderstood. Compare records from independent systems, verify timestamps and explain what the available evidence cannot establish.
The circle encompasses the complete 19-component reference model. It is not a census of every AI incident or a live view of your environment.
How to read the threat heat map
This map shows how often selected source cases identify each AI Harness component. Use it to decide where to investigate. The colors do not measure how dangerous a component is.
- Inner ring — research demonstrations: Three research cases show how an AI system may be exploited.
- Middle ring — reported operational failures: Two reported failures show how problems occurred in operating systems.
- Outer ring — lifecycle and control review: A retrospective analysis identifies components involved in control weaknesses across the system lifecycle.
Number in each cell: The count of distinct cases in that ring that identify the component. A component can appear in more than one ring.
Color: Darker shading means more case associations. It does not measure severity, probability, attack frequency, monitoring coverage or a vendor’s current exposure.
Missing, bypassed or ineffective control: A case may name a control because it failed or was absent. This does not mean the control was installed or working.
Zero: No matching case in this selected evidence. Zero does not mean the component is secure or has no risk.
Next action: Identify the component owner, exposed interfaces, independent controls and evidence of results. Prioritize actual control tests over interpreting colors as risk scores.
What is established
ServiceNow, GitHub MCP and ForcedLeak are presented as research demonstrations. Primary reports were checked for this release. Asana and Replit remain source-reported operational cases, not independently reverified incidents. No malicious actor is asserted for those two cases.
Component-to-case mappings and suggested controls are Richard E. Davis's analytical interpretation in the supplied reference. The lifecycle ring is explicitly retrospective analysis, not known observed attack activity.
MITRE ATLAS and OWASP identifiers are preserved as source-assigned taxonomy. This build does not independently certify each mapping or establish that a technique occurred. Vendor names are descriptive; no current configuration, partnership certification or deployed HoundBytes detection is asserted.
Numerical alternative: all 19 components, unfiltered corpus totals
Wide reference table: scroll horizontally to read every column. Keyboard users can focus the table and use the arrow keys.
| Component | Plane | Research | Operational | Lifecycle analysis |
|---|---|---|---|---|
| 6.00 Generative model | Control | 2 | 1 | 0 |
| 6.01 Inputs & external content | Boundary | 3 | 2 | 0 |
| 6.02 Orchestrator / agent loop | Control | 1 | 2 | 0 |
| 6.03 RAG subsystem | Data | 2 | 1 | 0 |
| 6.04 Memory & state | Data | 1 | 1 | 0 |
| 6.05 Context assembly | Control | 3 | 2 | 0 |
| 6.06 Input / output guardrails | Control | 1 | 0 | 0 |
| 6.07 Policy / authorization gate | Identity & action | 3 | 1 | 0 |
| 6.08 Credential broker | Identity & action | 2 | 2 | 0 |
| 6.09 Extensions / MCP / skills | Identity & action | 1 | 1 | 0 |
| 6.10 Tool dispatcher | Identity & action | 3 | 1 | 0 |
| 6.11 Sandbox / downstream systems | Data | 3 | 1 | 0 |
| 6.12 Observability & audit | Evidence | 3 | 2 | 0 |
| 6.13 Identity & delegation | Lifecycle | 0 | 0 | 3 |
| 6.14 Evaluation & conformance | Lifecycle | 0 | 0 | 3 |
| 6.15 Human oversight console | Lifecycle | 0 | 0 | 3 |
| 6.16 AI BOM & attestation | Lifecycle | 0 | 0 | 3 |
| 6.17 Authority & consumption | Lifecycle | 0 | 0 | 2 |
| 6.18 Transparency & recourse | Lifecycle | 0 | 0 | 1 |
Source basis: supplied davisAIgov / HoundBytes / Elastic Integrated Threat Map, reference v1.1, and AI Harness Telemetry to SIEM — Component Field Map, v1.0, 27 October 2026. Site release 14.1.1; interactive module v1.0.0. Historical reference-model versions remain identified as historical. Download the source-derived visual dataset.