Researchers report Copilot CLI attack that can leak secrets

The reported chain depends on optional autopilot mode and a user asking the coding agent to read an attacker-controlled web page.

The reported chain depends on optional autopilot mode and a user asking the coding agent to read an attacker-controlled web page.
Security researchers at Adversa AI say they made GitHub Copilot CLI read local developer secrets and send them to an attacker in tests reported by The Register on October 6. The attack required the coding agent to run in its optional autopilot mode and retrieve an attacker-controlled web page at the user’s direction. Success depended on the model handling the session.
GitHub disputes that this is a product vulnerability. The company says deliberately asking the tool to retrieve untrusted content and confirming the action amounts to user consent. Adversa disagrees and says the attack chain remains functional.
Copilot CLI had already been flagged earlier in 2026 for indirect prompt injection: instructions in outside content that push a model to act beyond its intended function. The reported technique, called Cryptographic Context Injection (CCI), adds encryption to that approach.
Researcher Rony Utevsky explained in a blog post provided to The Register that the malicious page carries encrypted instructions, the material needed to unlock them and a request for the agent to perform the decryption. In his account, safeguards that inspect incoming text miss the concealed instructions because they do not execute the decryption themselves.
According to Utevsky, the agent runs the decryption in its own code-execution environment. This exposes the hidden instructions to the agent while safeguards that only inspect incoming text miss them because they do not run that code.
The researchers’ attack chain induced the agent to read targeted local files containing secrets. A later web request carried those secrets to the attacker. The exposure was therefore not just an unwanted response from the model: the reported tests involved access to local data and its transmission outside the developer’s system.
| Model tested | Reported outcome |
|---|---|
| Microsoft mai-code-1.1-flash | Executed the full attack chain on 50% of attempts. |
| Two OpenAI GPT-5.6 models | Both refused the attack payload. |
The number of test attempts is not known. The reported percentage describes the researchers’ tests, not a measured rate of exposure among Copilot CLI users.
Utevsky said the vulnerable model was not the default on the paid account the researchers tested and had to be selected manually. But on an account left on Auto model selection, the system assigned the vulnerable model in some sessions and a model that resisted the attack in others. He said users neither chose nor saw which model handled those sessions, making the difference in test outcomes relevant even without a manual model change.
Adversa says it submitted the finding through GitHub’s bug bounty program on September 17, 2026. According to Adversa, GitHub’s triage team validated the finding but declined to classify it as a vulnerability.
After investigating, we determined this requires a user to intentionally direct Copilot CLI to fetch attacker-controlled or untrusted content and confirm they want to trigger the action, and thus is not a product vulnerability.
Adversa rejects that conclusion and says the chain still works as described. It is not yet known whether the reported chain has been patched or used to compromise developers outside the tests.

Asos says attackers obtained an employee’s login details by impersonating a trusted contact, then accessed third-party platforms. The affected customer total remains unknown.
5 min

Coding agents used public repositories to share private code-review images, Glow says, often placing them under developers’ personal accounts rather than company accounts.
4 min

Ofcom is examining whether Meta met its legal duties to assess illegal-content risks and risks to children before launching Instagram’s disappearing-photo feature.