New Technique Medium TD-2026-0105 For AppSec, AI Eng, Red Team AI-assisted
Incomplete-fix bypass in PraisonAI tool loading: sibling tool_override path still violates invariant at 4.6.31
A benign, control-gated probe study compared original-trigger replay with a sibling-path probe across LLM agent frameworks. At PraisonAI 4.6.31, the original tool-loading trigger was blocked while the tool_override path still violated the invariant, reproducing the publicly documented CVE-2026-44334 bypass of CVE-2026-40287. Neither tested path showed a violation at 4.6.32.
Why it matters
Patching only the original proof-of-concept trigger can leave equivalent code paths exploitable, so organizations running affected PraisonAI releases may remain exposed after the first fix. The study reports the probes invoke loader functions directly, so it does not establish a full end-to-end exploit.
Do this
- Confirm PraisonAI is at 4.6.32 or later in all environments that load tools dynamically.
- Review tool_override handling and related loader code paths, not only the original tool-loading trigger, when validating a fix.
- Re-test sibling code paths after patching to confirm the security invariant holds beyond the original reproduction.
- For CrewAI, confirm deployments are on 1.11.1 or later, since tested release 1.11.0 showed a violation.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“At PraisonAI 4.6.31, the original tool-loading trigger is blocked while the tool_override path still violates the invariant.”
“This reproduces the publicly documented CVE-2026-44334 bypass of CVE-2026-40287; neither tested path shows a violation at 4.6.32.”
“For CrewAI, we observe a violation in tested release 1.11.0 and confinement in 1.11.1, without claiming the earliest fixed release.”
Sources
Affected / related
PraisonAI, CrewAI, Langroid, Agno, Incomplete-fix patch bypass validation
Published 2026-11-16 04:39 UTC · updated 2026-11-17 08:35 UTC · AI-assisted summary, checked against quoted sources · How we score