Disclosed weakness · no CVE recorded

Pydantic Harness JWT redaction could block concurrent agent work

Pydantic · pydantic-ai-harness · secret-redaction guardrails

First documented disclosure · Review checkpoint

Summary

Crafted text scanned by the JWT redaction pattern could occupy the event loop and delay other requests and agent runs.

Severity and exploitation evidence

High · 7.5/10 · CVSS 3.1. Severity source.

Exploitation not established: The reviewed sources establish the weakness or a controlled demonstration; malicious exploitation against users is not established.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The JWT secret-detection regular expression backtracked in quadratic time while synchronous guards ran on the event loop.

Conditions required

An application uses redact_secrets or secret_data with the default JWT pattern on attacker-influenced prompts, tool results or model output.

Potential impact

Denial of service affecting concurrent work in the same process.

Affected versions and configurations

pydantic-ai-harness >= 0.16.0, <= 0.69.0.

Fix and mitigation status

pydantic-ai-harness 0.70.0.

Practical next steps

Upgrade the harness or omit the JWT pattern using an explicit secret_data(only=[...]) pattern list.

Operating lesson

Bound the cost of content inspection before allowing it to block shared execution.

Evidence and limits

November 18 is the original advisory date; this is a backfill. The maintainer assigns High/CVSS 3.1 7.5 and no CVE is shown. This JWT guardrail issue is distinct from CVE-2026-133161, which concerns ANSI stripping.

Sources