Skip to content
AI Security

CamoLeak: a PR comment made Copilot leak a private repo

CVE-2025-59145 (CVSS 9.6) turned GitHub Copilot into an exfiltration channel: a hidden instruction in a PR description made the agent leak code and secrets from a private repo, zero-click. What it teaches about trusting agents.

The attack that redefined agent trust

In October 2025 CamoLeak went public (CVE-2025-59145, CVSS 9.6), one of the most elegant and frightening flaws of the year. An attacker hid an instruction in a pull request description. When the victim opened GitHub Copilot Chat in that context, the agent read the hidden instruction as a legitimate order and started exfiltrating code and secrets from the private repository, with no click from the victim. Zero-click.

The clever detail: exfiltration went out through GitHub's own image proxy (Camo). The agent encoded the stolen data character by character into a pre-generated dictionary of Camo image URLs, and each image request carried a piece of the secret to an attacker server. AWS keys, tokens, even the description of a zero-day stored in a private issue. GitHub fixed it on August 14, 2025, by disabling image rendering in Copilot Chat.

Why it matters for your SaaS

You may not use Copilot Chat on a private repo, but CamoLeak is a portrait of a problem that applies to ANY AI agent you connect to your product: the data channel becomes the command channel. If the agent reads content a stranger controls (a PR description, a support ticket, a public page, an email), that content can carry an instruction the model obeys.

In AI-built SaaS, it shows up in obvious places once you look:

  • A support agent that reads the customer's ticket and has a tool to change the account.
  • An agent that summarizes a user-uploaded document and then calls an internal API.
  • A workflow that processes inbound email and triggers an action.

In all of them, external (untrusted) content and action permission (trusted) meet. That is exactly where injection turns into damage.

The defense is not trusting the model

CamoLeak's lesson is not "stop using agents". It is treating all external content as untrusted and separating the data channel from the command channel:

  • What the agent READS (external context) should never become what the agent EXECUTES without a boundary.
  • A destructive action or one that touches sensitive data needs human approval or a strict allowlist.
  • Secrets out of the agent context: if the model cannot see the key, it cannot leak it.
  • A restricted output channel: CamoLeak abused the image proxy; limit where the agent can send data.

Where RET fits

When your SaaS connects an agent to login, billing, customer data or a state-changing tool, the permission boundary becomes an attack surface. RET's Risk Review reads exactly those flows, what the agent can read, what it can execute, with which approval and which output channel, and flags where external context can become a command before it becomes an incident.

Any agent that reads content a stranger controls can be hijacked by it. The defense is a permission boundary, not faith in the model.

Sources

Tags:#CamoLeak#GitHub Copilot#prompt injection#CVE-2025-59145#exfiltration
Gabriel Ferreira
Gabriel Ferreira

Founder & Lead Pentester

Gabriel Ferreira reviews production SaaS with focus on login, billing, data, and evidence founders can act on.

How much does it cost to test your system?

Answer four questions and see the package, price and timeline for your case, no sign-up. We refund 50% if the test finds no flaw.

See price and timeline for my case
Keep reading
View all