agent-trust-telemetry
Trust telemetry middleware for inter-agent communication — makes instruction contamination observable across traces.
Why
In agent orchestration, a trusted sender does not imply trusted content.
Tool descriptions from MCP servers, messages relayed via sampling, forwarded content from upstream agents — all of these can carry contaminated instructions even when the sender is legitimate. Single-message input filtering cannot capture this chain contamination.
Human → Agent A (contaminated) → Agent B → Agent C
↑
Passes through because sender is trusted
But the content is contaminated
agent-trust-telemetry makes this visible by adding security semantics over traces to your existing observability stack.
What It Does
- Normalized message envelope with provenance, execution phase, and content hash
- Policy violation taxonomy — 5 violation classes + 2 anomaly indicators
- One-hop risk inheritance — upstream contamination score propagation
- OTel-compatible evidence export — integrates with your existing tracing infrastructure
What It Does NOT Do
- Guarantee complete prompt injection defense
- Provide model-internal interpretability
- Prove agent capability authenticity
- Automate legal liability boundaries
3-Minute Repro (LangGraph)
git clone https://github.com/wharfe/agent-trust-telemetry.git
cd agent-trust-telemetry
pip install -e ".[dev,langgraph]"
python examples/langgraph/multi_agent_contamination.py
Expected behavior:
agent_ais flagged withaction=quarantine(poisoned instruction detected)- downstream nodes inherit contamination context and remain quarantined
- ancestor chain is preserved for traceability across hops
LangGraph Minimal Integration
from att.integrations.langgraph import TrustTelemetryCallback
callback = TrustTelemetryCallback()
result = graph.invoke(
{"messages": messages},
config={"callbacks": [callback]},
)
CLI Quick Start
pip install agent-trust-telemetry
# Evaluate a single message
att evaluate --message message.json
# Evaluate a JSONL stream
att evaluate --stream messages.jsonl
# View evaluation report
att report --input evaluations.jsonl --format table
Output Example
{
"risk_score": 82,
"severity": "high",
"policy_classes": [
{"name": "instruction_override", "confidence": 0.91, "severity": "high"}
],
"anomaly_indicators": [
{"name": "provenance_or_metadata_drift", "subclass": "parent_flagged_propagation", "confidence": 0.70, "severity": "high"}
],
"evidence": [
"override phrase detected near end of long context",
"parent message msg:xxx was flagged (risk_score: 65, action: warn)"
],
"recommended_action": "quarantine"
}
quarantine does not simply block a single message — it pauses propagation to downstream agents.
How It Relates to Existing Tools
| Tool | Focus |
|---|---|
| Lakera Guard | Point-in-time content screening |
| LangSmith | Observability / debugging / tracing |
| This tool | Security semantics over traces + inter-agent trust propagation |
Not a competitor — a complement. Designed to layer on top of existing OTel stacks without requiring a proprietary monitoring platform.
Architecture
Layer 1 Cheap Deterministic Known markers, pattern matching
Layer 2 Contextual Classifier Maps to policy violation classes (LLM used only for refinement)
Layer 3 Risk Aggregation Upstream inheritance, session accumulation, role-transition drift
MVP implements Layer 1 only. The core evaluation engine has zero external LLM API dependencies.
Status
Alpha — MVP complete
- Message Envelope Schema (JSON Schema Draft 2020-12)
- Policy Violation Taxonomy v0.1
- Architectural decisions documented (execution phases, propagation, tool metadata)
- JSON Schema validation (envelope + output contract)
- Layer 1 evaluation engine (regex pattern matching)
- One-hop risk inheritance
- Tool metadata drift detection
- CLI (
att evaluate/att report/att quarantine) - OTel span/event export
- End-to-end demo (Docker Compose + Jaeger)
Contributing
See CONTRIBUTING.md for development setup and guidelines.
Security
To report a vulnerability privately, use GitHub Security Advisories:
For non-sensitive security discussion, open an issue:
License
Metadata
Release files for agent-trust-telemetry 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| agent_trust_telemetry-0.1.0.tar.gz | 29.2 kB | Without provenance | Details for agent_trust_telemetry-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| agent_trust_telemetry-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for agent_trust_telemetry-0.1.0-py3-none-any.whl |
Total release size: 66.5 kB
Release files / agent_trust_telemetry-0.1.0.tar.gz
| Download URL | agent_trust_telemetry-0.1.0.tar.gz |
|---|---|
| Size | 29.2 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
7769541d9be5520dee2d5861d43cecde50dba89cce9f6ba403623b3a9f18a6c6
|
|
BLAKE2b-256 checksum How to use checksums |
271ddf4c27eb7a9949bdcb5530f1ee3bef60a72a29e305354b9a489087f3992e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.8.10
|
Release files / agent_trust_telemetry-0.1.0-py3-none-any.whl
| Download URL | agent_trust_telemetry-0.1.0-py3-none-any.whl |
|---|---|
| Size | 37.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
2012823f53002e89fe082954ca51a401342d7420a541a22cbc4756b97183f53d
|
|
BLAKE2b-256 checksum How to use checksums |
e1e61baf29799b56881198540ff04f1cae137d9ea1d69a4800d5a1458cc2719a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.8.10
|
Security for agent-trust-telemetry 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.