Scope: what this does not show
Specific product versions and researcher-controlled demonstrations. Earlier coding-agent disclosures concern omitted command content or approval-free configuration changes; Salt’s Manus case concerns warning timing. These are not studies of how users interpret approval prompts.
Corroborated: Supported by at least two independent sources.
Evidence
Jul 28, 2025
Tracebit shows Gemini CLI could silently run attacker commands when reading untrusted code
The approval display omitted parts of the command that actually ran.
Aug 1, 2025
CurXecute: prompt injection could make Cursor create MCP config and run commands (CVE-2025-54135)
Creating a new MCP configuration file needed no approval, although editing one did.
Oct 1, 2026
Salt Labs reports a resolved Manus flaw that ran injected code before warning
Salt reports Manus displaying an approval warning after untrusted code had already executed; it states the flaw has been resolved. This adds a non-coding workflow to the earlier coding-agent cases.
Key questions that rely on this finding
- Where are deployed AI agents actually being exploited?Exploitation often reaches deployed agents through connectors, credentials, tools and packages.
Status history
- 2025-07-28ReportedGemini CLI's approval display omitted parts of executed commands. · record
- 2025-08-01CorroboratedCursor did not require approval for creating sensitive files. · record
- 2026-10-13CorroboratedreconfirmedSalt Labs adds an email-agent demonstration in which untrusted code executed before the warning requested approval. This broadens the finding beyond coding agents and adds warning timing to the earlier omissions and approval-free changes. · record