// Security Advisory Triage

Pydantic AI web chat UI allows remote sites to trigger agent tool execution

GitHub advisories — pip · published 2026-10-09 13:21 UTC
ACTION NEEDED PATCH_NOW SECURITYAPI
WHAT CHANGEDCVE-2026-107295 published 8 October 2026. The web chat UI exposed by Agent.to_web() and clai web in pydantic-ai and pydantic-ai-slim allows a website visited by the developer to trigger agent runs and execute server-side tools on the local chat endpoint. Patched in versions 1.107.4 and 2.28.0.
WHY IT MATTERSIf you run the local web UI during development, a malicious site in another browser tab can execute arbitrary agent tools on your machine without interaction. This is server-side execution, not just data exfiltration — any tool the agent has access to can be invoked remotely.
WHAT TO DOUpdate pydantic-ai or pydantic-ai-slim to 1.107.4 (1.x branch) or 2.28.0 (2.x branch). If you do not use Agent.to_web() or clai web, verify your dependency tree does not pull in the vulnerable versions through another package.
Model-reported confidence 95% (not measured accuracy) Primary source ↗ written by anthropic/claude-sonnet-4.5
Show the reasoning — raw diff, triage verdict, confidence
RAW DIFF
+ghsa_id: GHSA-h4xc-3qfq-jf93
+summary: Pydantic AI Web chat UI (`Agent.to_web()`, `clai web`): a website visited by the developer can trigger agent runs and server-side tool execution on the local chat endpoint
+severity: high
+published_at: 2026-10-08T17:16:36Z
+cve_id: CVE-2026-107295
+vulnerabilities[].package.name: pydantic-ai, pydantic-ai, pydantic-ai-slim, pydantic-ai-slim
+vulnerabilities[].first_patched_version: 1.107.4, 2.28.0, 1.107.4, 2.28.0
HOW IT WAS JUDGED
Triage signals below are model-reported assertions, preserved verbatim; they are not independent arithmetic verification.
triage verdictMATERIAL at 0.92, needed 0.5
triage reasoningPydantic AI SDK vulnerability enables arbitrary tool execution via web UI; affects AI agent deployments.
signalsAI SDK vulnerability · server-side tool execution · local endpoint compromise · high severity · patched versions available · pydantic-ai in dependency tree of AI services
suppression—
triage modelanthropic/claude-haiku-4.5 ($0.00000)
writer modelanthropic/claude-sonnet-4.5 ($0.00000)