CVE-2026-47391 is a remote code execution vulnerability in PraisonAI prior to version 4.6.40 affecting the first-party Agent-to-Agent (A2A) server example. The vulnerable example exposes an unauthenticated A2A JSON-RPC endpoint, listens on all interfaces, and registers a calculate(expression) tool implemented with Python eval(). An unauthenticated remote attacker can submit a message/send request to the A2A endpoint, causing the request to reach agent.chat(); when the backing LLM invokes the registered calculate tool on attacker-influenced input, arbitrary Python code can execute in the server process. The exposed unauthenticated A2A surface also includes task history and task cancellation functionality, expanding the confidentiality and integrity implications beyond code execution. The issue affects deployments using the official A2A example or similar public unauthenticated A2A deployments that expose unsafe tools.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
No public exploits tracked yet. Mallory keeps watching.
No public exploit code observed for this vulnerability.
4 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
An unauthenticated remote code execution vulnerability in PraisonAI's first-party A2A server example caused by exposing an unauthenticated JSON-RPC endpoint and an unsafe `calculate(expression)` tool implemented with Python `eval()`.
A PraisonAI A2A server example vulnerability exposing an unauthenticated calculate(expression) tool backed by Python eval(), which was exploited to induce the victim's own LLM agent to execute a reverse shell.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.