CVE-2026-61536 (GHSA-64vx-6h2c-rjh7): banks resolves an attacker-controlled tool import path without an allowlist, so a model's tool call can run arbitrary code
The GitHub-reviewed advisory GHSA-64vx-6h2c-rjh7 says that the banks package resolves the import_path of Tool JavaScript Object Notation (JSON) objects in rendered completion blocks through importlib with no allowlist or sanitization, so when the large language model (LLM) emits a tool call matching an attacker-supplied tool name, a callable such as os.system is invoked with the model's arguments, giving arbitrary code execution in the hosting process.
The record
- Kind
- Vulnerability
- Source record
- GHSA-64vx-6h2c-rjh7: banks: Banks: Unsafe importlib.import_module of attacker-controlled Tool.import_path in CompletionExtension allows RCE (GitHub-reviewed advisory via OSV)
- Record published
- 2026-11-05
- Posted here
- 2026-11-06
- World
- Software agents · Enterprise work agents
- Controls concerned
- Containment and isolation C-E5; Pre-execution validation C-D1
- Checked
- Checked against its record on 2026-11-06 by a reader who did not draft it; report
verification/reports/SIG-2026-11-06-003-2026-11-06.md.
A signal's controls are the ones its record points at, the exposure; a case's controls are the ones its records show were absent or failed, so the two can differ.
Corrections
Changes to what this signal says, with the reason. Every earlier version is kept.
- 2026-11-06: Summary. The record calls banks a package, not a "prompt library" (beyond the record); JSON was not spelled out at first use.
Show the change
Before: The GitHub-reviewed advisory GHSA-64vx-6h2c-rjh7 says that the banks prompt library resolves the import_path of Tool JSON in rendered completion blocks through importlib with no allowlist or sanitization, so when the large language model (LLM) emits a tool call matching an attacker-supplied tool name, a callable such as os.system is invoked with the model's arguments, giving arbitrary code execution in the hosting process.After: The GitHub-reviewed advisory GHSA-64vx-6h2c-rjh7 says that the banks package resolves the import_path of Tool JavaScript Object Notation (JSON) objects in rendered completion blocks through importlib with no allowlist or sanitization, so when the large language model (LLM) emits a tool call matching an attacker-supplied tool name, a callable such as os.system is invoked with the model's arguments, giving arbitrary code execution in the hosting process.
What a signal is. A signal records what a primary record says; it assigns no fault. It is lighter than a case: one dated item from one record, tagged with the world, sector and controls it touches. It becomes a case when it meets the bar for one.
New signals like this, once a week. A short email when a case, signal or rule touches software agents. Free; unsubscribe in one click.
Get the weekly emailNewer: CVE-2026-54140: OpenHuman passes unneutralized tool descriptions from external Model Context Protocol (MCP) servers into the agent's context, so a malicious server can steer the agent to invoke other tools (fixed in 0.57.13)
Older: CVE-2026-113663: AG2's sandbox shell tool filter misses newline separators, so an attacker-influenced agent command can run an extra arbitrary statement, which runs on the host under the default backend (fixed in ag2 1.1.1)