Failure pattern

Open exfiltration channel

Model output could carry data to an outside party, for example through a link or an image address.

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-2025-0004
Microsoft
Separate untrusted inbound content from internal data in retrieval, and restrict the URLs an assistant’s output may reference.
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.

4 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-2025-0004
    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
  4. 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