Failure pattern

Indirect prompt injection

Content the agent read from an untrusted source was treated as instructions.

Preventive controls from the record

IncidentControl
INC-2025-0018
GitHub
Strip or expose hidden content before it reaches the assistant, and do not render remote images in assistant output.
INC-2025-0016
Salesforce
Treat form input as untrusted in agent context, and audit allow-listed domains for expiry and ownership.
INC-2025-0013
Perplexity
Keep page content separate from user instructions, and require confirmation before the agent acts on another site.
INC-2025-0012
Cursor (Anysphere)
Treat files that configure execution as protected: require approval to create them, and never run them before approval.
INC-2025-0010
Google
Match allow-listed commands exactly, show the full command before it runs, and run agents on untrusted code in a sandbox.
INC-2025-0006
Supabase
Give agents read-only, project-scoped database access and never a key that bypasses access rules.
INC-2025-0004
Microsoft
Separate untrusted inbound content from internal data in retrieval, and restrict the URLs an assistant’s output may reference.
INC-2025-0003
GitHub
Scope the agent’s token to the repositories the task needs, and keep one repository per session.
INC-2024-0004
Slack (Salesforce)
Keep retrieved content from untrusted authors out of answers about private data, and do not render model-built links that can carry data out.

Controls are editorial summaries by Incidents.si, not statements by the organizations involved.

9 incidents with this pattern

  1. INC-2025-0018
    CamoLeak: hidden pull request text could make Copilot Chat leak private code

    Legit Security showed that a hidden comment in a pull request could instruct Copilot Chat to leak private repository content through GitHub’s own image proxy.

    GitHub · Possible exfiltration of private source code and secrets; no exploitation reported

    SEV2potentialPrompt injectionDemonstrated
  2. INC-2025-0016
    ForcedLeak: a web form could make Salesforce Agentforce send out CRM data

    Noma Security showed that text submitted through a public lead form could instruct Agentforce to send CRM data to an expired domain that Salesforce’s policy still trusted.

    Salesforce · Possible exfiltration of CRM data; no exploitation in the wild reported

    SEV2potentialPrompt injectionDemonstrated
  3. INC-2025-0013
    Hidden webpage text makes Perplexity’s Comet browser leak a login code

    Brave showed that instructions hidden in a Reddit comment could make Comet, while summarizing the page, read the user’s email and send a one-time login code to an attacker.

    Perplexity · Demonstrated account takeover path through the user’s email; no exploitation reported

    SEV2potentialPrompt injectionDemonstrated
  4. INC-2025-0012
    CurXecute: prompt injection leads to code execution in the Cursor editor

    Aim Security found that injected text could make Cursor’s agent write an MCP configuration file that ran commands before the user could reject the edit. Fixed as CVE-2025-54135.

    Cursor (Anysphere) · Possible arbitrary code execution on a developer machine; no exploitation in the wild reported

    SEV2potentialPrompt injectionDemonstrated
  5. INC-2025-0010
    Gemini CLI could be made to run hidden commands from a repository file

    Tracebit showed that a prompt injection hidden in a repository’s README could make Gemini CLI run a command the user never saw. Google fixed it in version 0.1.14.

    Google · Demonstrated exfiltration of environment variables from a developer machine; no exploitation reported

    SEV3potentialPrompt injectionDemonstrated
  6. INC-2025-0006
    Support ticket text steers a coding agent into leaking private database tables

    General Analysis showed that instructions hidden in a support ticket could make a coding agent with a privileged Supabase key read a private table and write its contents back into the ticket.

    Supabase · Demonstrated leak of a private table on dummy data; no customer leak reported

    SEV3potentialPrompt injectionDemonstrated
  7. INC-2025-0004
    EchoLeak: an email alone could make Microsoft 365 Copilot leak data

    Aim Security found that a crafted email could cause Microsoft 365 Copilot to send data from a user’s context to an attacker with no click. Microsoft fixed it as CVE-2025-32711.

    Microsoft · Possible exfiltration of data from Copilot’s context without any user action; no exploitation in the wild reported

    SEV2potentialPrompt injectionDemonstrated
  8. INC-2025-0003
    Malicious GitHub issue steers agents into leaking private repository data

    Invariant Labs showed that an issue planted in a public repository could make an agent using the GitHub MCP server copy data from the user’s private repositories into a public pull request.

    GitHub · Demonstrated leak of private repository details into a public pull request; no exploitation reported

    SEV2potentialPrompt injectionDemonstrated
  9. INC-2024-0004
    Researchers show Slack AI can be tricked into leaking private channel data

    PromptArmor showed that a message in a public Slack channel could steer Slack AI into revealing data from a private channel the attacker could not access.

    Slack (Salesforce) · Demonstrated exfiltration of a secret from a private channel; no exploitation reported

    SEV2potentialPrompt injectionDemonstrated