CVE-2026-76072

Affects: prompt injection

CVSS
HIGH · 7.4v3.1
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:H/A:H
Published
2026-08-24
Weakness
CWE-184
Source
nvd.nist.gov/vuln/detail/CVE-2026-76072

Description

The Continue CLI applies an incomplete denylist as its only barrier to destructive shell commands when running unattended. In headless mode and auto mode the default policy in extensions/cli/src/permissions/defaultPolicies.ts grants the Bash tool the allow permission, and permissionChecker.ts hard-blocks a command only when the terminal-security evaluator returns a disabled verdict, so isCriticalCommand in packages/terminal-security/src/evaluateTerminalCommandSecurity.ts is the sole control. Its dangerous-path test matches only /, /, ~, ~/, /usr, /etc, /bin and /sbin and their prefixes, so a recursive forced removal of /home, /root, /var, /opt or /srv is not disabled. The command line is parsed with shell-quote, which reduces $HOME to an empty token, so rm -rf $HOME also fails the dangerous-path test while the shell re-expands the variable when the command is spawned. find with -delete is rated high risk rather than disabled, and shred, wipefs, truncate and pkexec are not handled. Because the agent autonomously reads content it does not control, including fetched web pages, repository files and issue text, an indirect prompt injection in that content can cause an unattended run to destroy the invoking user’s data.

References

How GTK Cyber trains on this

AI security training at GTK Cyber covers the LLM and ML-pipeline vulnerability classes that vulnerabilities like CVE-2026-76072 fall into. Our hands-on courses are taught by Charles Givre and other practitioners who break and defend production AI systems.

AI Red-Teaming course →·Browse MITRE ATLAS techniques

Related AI/LLM CVEs

AI security training, taught by people who do the work.

Hands-on courses on adversarial AI, prompt injection, and ML pipeline security.

Explore AI Red-Teaming