Pydantic AI v2.70.0 fixes two high-severity vulnerabilities in Harness; malicious input can block the event loop for up to several minutes
v2.70.0 (2026-11-17)
Pydantic AI v2.70.0 fixes two high-severity vulnerabilities in Pydantic AI Harness; the fixed version is pydantic-ai-harness 0.70.0. GHSA-m4cr-m4qc-9hh3: when strip_ansi=True, the pattern in ToolOutputLimits that removes ANSI escape sequences has time cost that grows quadratically with line length; tool output that an attacker can influence can contain hundreds of KB of unterminated escape sequences, which can occupy the event loop for tens of seconds to several minutes. GHSA-5m5q-5qj4-gx9h: backtracking in the jwt mode of secret redaction has time cost that grows quadratically with text length; processing attacker-influenceable text with this mode of redact_secrets or secret_data() (included by default) is affected; after the fix, JWTs whose header segment contains -eyJ are no longer matched. The version also includes new features such as StreamGuardrail streaming segment-by-segment screening and multiple fixes.