Pydantic released pydantic-ai v1.107.9 (and v2.59.0) fixing a high-severity security issue in the web_fetch tool's HTML conversion. Fetching attacker-controlled HTML with many elements could delay unrelated work in the same process. Users who let the tool fetch untrusted pages are advised to upgrade.
Pydantic AI released v2.52.0 on 2026-09-29, fixing a moderate security issue (GHSA-v36g-jcw9-x7cw) in the local web_fetch tool, with the fix backported to v1.107.7. The release also moves pydantic-ai-harness into the main repository and ships the first pydantic-clai2 release. Users on affected versions should upgrade.
Pydantic AI released v2.52.0 on 2026-09-29, fixing a moderate security issue (GHSA-v36g-jcw9-x7cw) in the local web_fetch tool, with the fix backported to v1.107.7. The release also moves pydantic-ai-harness into the main repository and ships the first pydantic-clai2 release. Users on affected versions should upgrade.
Pydantic released pydantic-ai v2.61.0 on 2026-10-23, a broad update spanning features, compatibility changes, and bug fixes. Notable additions include structured citations on TextPart with normalization for Bedrock, OpenRouter, and xAI, model names for kimi-k3.1-pro, and a public extension surface for the durable engine.
Pydantic released pydantic-ai v2.61.0 on 2026-10-23, a broad update spanning features, compatibility changes, and bug fixes. Notable additions include structured citations on TextPart with normalization for Bedrock, OpenRouter, and xAI, model names for kimi-k3.1-pro, and a public extension surface for the durable engine.
Pydantic AI v1.107.2 and v2.24.0 fix an availability vulnerability where the local web_fetch tool or FileUrl media downloads could exhaust process memory by downloading unbounded remote content. The patch enforces a default 50 MiB download cap to prevent worker crashes.
Pydantic AI v1.107.2 and v2.24.0 fix an availability vulnerability where the local web_fetch tool or FileUrl media downloads could exhaust process memory by downloading unbounded remote content. The patch enforces a default 50 MiB download cap to prevent worker crashes.
Pydantic AI shipped v2.60.0, a large release adding sandbox harness capabilities (Docker, BoxLite micro-VMs), a judge-ranked RelevanceCompaction history mechanism, and SmartSearch code search. It also deprecates RetryPromptPart and OpenAIDecisionsProvider, hardens toolset security, and fixes dozens of realtime, Bedrock, and durability bugs.
Pydantic AI shipped v2.60.0, a large release adding sandbox harness capabilities (Docker, BoxLite micro-VMs), a judge-ranked RelevanceCompaction history mechanism, and SmartSearch code search. It also deprecates RetryPromptPart and OpenAIDecisionsProvider, hardens toolset security, and fixes dozens of realtime, Bedrock, and durability bugs.
pydantic/pydantic-ai v2.28.0 patches a high-severity security flaw in the development web chat UI (Agent.to_web(), clai web). The chat endpoint lacked content-type validation, allowing cross-origin requests to trigger agent execution with local privileges. The fix requires Content-Type: application/json, rejecting other requests before parsing.
pydantic/pydantic-ai v2.28.0 patches a high-severity security flaw in the development web chat UI (Agent.to_web(), clai web). The chat endpoint lacked content-type validation, allowing cross-origin requests to trigger agent execution with local privileges. The fix requires Content-Type: application/json, rejecting other requests before parsing.