Loading vulnerability…
CVE-2026-75858 is an improper code-generation control vulnerability in CodeWhale codewhale and codewhale-tui versions 0.8.41 through 0.8.63. The rlm_eval tool declares ApprovalRequirement::Auto, which causes the tool engine to suppress approval prompts rather than enforce the configured --approval-policy. Consequently, model-supplied Python is executed through a local Python interpreter without explicit user approval or an audit step. Reporting also identifies the interactive shell-execution tool as affected by the same automatic-approval design. Indirect prompt injection in agent-readable untrusted content can cause the model to invoke an affected tool with attacker-controlled code.
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.
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
Vendor-confirmed product mapping. Mallory continuously reconciles this list against your asset inventory.
3 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
An authorization-bypass/code-execution vulnerability in CodeWhale interactive execution tools. The rlm_eval and exec_shell_interact tools designated themselves as auto-approved, allowing indirect prompt injection to trigger unprompted Python or shell command execution under the local developer's account.
An approval-policy bypass in CodeWhale's rlm_eval tool that permits an LLM-controlled Python string to run unsandboxed on the host with the user's privileges, without an approval prompt. The companion rlm_open tool can stage attacker-controlled content for execution.
A remote code execution vulnerability in CodeWhale's rlm_eval tool that can allow arbitrary model-supplied Python code to execute without user approval when the agent processes untrusted content via prompt injection.
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.