Agent incident database
Slack AI could be steered into leaking secrets from private channels
20 Aug 2024Salesforce (Slack) · Slack AIdata exfiltrationdisclosed vulnerability
PromptArmor showed that a message posted in a public Slack channel could make Slack AI put a user's private API key into a link. If the user clicked it, the key went to the attacker.
How it unfolded
Each step on its own looked like ordinary agent work. The harm came from the sequence.
- 1
Read a message an attacker posted in a public channel
- 2
Read the user's API key from a private channel to answer their question
- 3
Model followed the attacker's instruction and built a link containing the key
- 4
Showed the user a link that sends the key to the attacker's server when clicked
harm
Which policy would have stopped this?
In our reading of the reports, these policies from the open library stop the harmful step before it runs.
- Restrict outbound to approved domains: The answer contained a link to an address outside the approved domains.
- Block external send after sensitive data read: The key was read from a private channel first, then sent out in the link.
Open Restrict outbound to approved domains in the playground and build the path yourself.
This is our reading of public reporting. We haven't run it against the original systems.
What happened
PromptArmor published the research on 20 August 2024 after disclosing it to Slack on 14 August. A user keeps an API key in a private channel only they can see. An attacker creates a public channel and posts an instruction telling Slack AI to add the key as a parameter to a link labelled as a re-authentication prompt. When the user asks Slack AI for their key, both messages land in the same context; Slack AI follows the attacker's instruction and shows the link with the key embedded, while its citations do not show the attacker's message. Clicking the link sends the key to the attacker's server. PromptArmor noted that Slack AI's ingestion of uploaded files, added on 14 August, widened the ways such instructions could arrive.
Outcome
Slack initially told PromptArmor the behaviour was intended. On 21 August 2024 a Salesforce spokesperson said it had deployed a patch and had no evidence of unauthorised access to customer data.
Sources
- Data Exfiltration from Slack AI via indirect prompt injection, PromptArmor, 2024-08-20
- Slack AI can leak private data via prompt injection, The Register, 2024-08-21
A summary of public reporting. The step-by-step timeline is our reading of those sources; where they differ, we say so above.
Your agents
Could this happen to yours?
For developers
Build a path yourself
Recreate a sequence in the playground and see how policies treat it.