Sophos says a fake interview assignment compromised an employee’s Mac through a trusted Visual Studio Code workspace. Its endpoint protection was installed but broken. During containment, a separate application authorization survived the password reset and session revocation.

Chief information security officer Ross McKerchar’s November 4 account revisits a September incident. The detailed root-cause analysis, INC-2026-002, was published October 20. This is a recent explanation of an older compromise, not a newly discovered September attack.

Evidence status: actual compromise reported by the affected vendor. Sophos reports one affected internal endpoint and unauthorized reads from two internal repositories. It says no product, service or customer environment was accessed.

Trusting the workspace enabled the first command

The employee cloned the assignment and opened it as a trusted workspace on September 14. Repository settings replaced the editor’s integrated terminal profile. Opening a terminal then launched the malicious command, leading to the BeaverTail loader and InvisibleFerret backdoor.

Sophos explicitly says this used a legitimate editor feature, with no software vulnerability exploited. Later, the employee granted VS Code Full Disk Access, expanding what its child processes could reach.

The backdoor could steal credentials, but capability is separate from observed theft. Sophos found no cloud keys, access tokens or passwords in the infected process environment. It assessed the remaining exposure there as the user’s SSH-agent session; the employee said only personal SSH keys were involved and revoked them.

For unfamiliar assignments, keep the folder in VS Code’s Restricted Mode while reviewing editor settings, tasks and setup scripts. Review inherited trust too: trusting a parent directory trusts its children. Restricted Mode limits automatic execution, but cannot restrain a malicious extension that ignores it.

If execution is necessary, use an isolated, disposable environment without corporate sessions, forwarded SSH access or production credentials. Those are containment recommendations drawn from the mechanism, rather than a claim that one editor setting guarantees safety.

Installed protection needs fresh evidence that it works

The Mac had last reported to Sophos Central on June 11. Its protection services were stopped, so it received no updates, enforced no protection and supplied no management telemetry.

An external partner reported the compromise on September 15. Sophos subsequently found other devices that passed identity and device-management checks while failing to report to its corporate endpoint console.

McKerchar recommends comparing devices still signing in with devices actually reporting security telemetry. Require recent, correctly registered endpoint health, not merely an installation record. Sophos says a functioning agent would have blocked the first payload; that is its own product team’s assessment.

Revoke app grants separately from sign-in sessions

A third-party OAuth authorization, permission granted to another application, survived the initial containment steps. On September 15 it made 27 unauthorized read requests against two internal source repositories, according to Sophos. The company revoked it on September 16 and observed no further use.

For a compromised developer identity, review and revoke affected application grants alongside passwords, sessions and tokens. Verify revocation in the source platform’s audit trail. A password reset alone did not terminate this authorization.

Sophos found no persistence mechanism on the Mac. Rebooting ended the in-memory implant, but the response also included evidence collection, access removal and device replacement. Rebooting alone would have left the application grant intact.

Caveat: Sophos both suffered the incident and sells the protection it evaluates; its detailed analysis leaves the agent failure’s engineering cause unresolved and makes no named-group attribution.