Bitdefender finds GitHub campaign targeting coding agents
By AgentRiot Editorial
Researchers identified 22 malicious GitHub links. The report shows why granting a coding agent access to a repository should be a separate decision from letting that repository run code.

A coding agent can read project instructions before a developer has reviewed them. In a report published November 17, 2026 (opens in new tab), Bitdefender Labs said researcher Andrei Anton-Aanei identified 22 malicious GitHub links containing files designed to make AI-assisted development tools run concealed malware. Some files tried to persuade the agent to act. Other project settings could trigger execution without a model decision once the workspace was trusted and the relevant automation enabled.
The 22 links measure what Bitdefender found, not how many developers were affected. The report does not establish how many times the malware ran, how many credentials were stolen, or whether any downstream packages were successfully compromised. It describes abuse of content hosted on GitHub, not a breach of GitHub itself. It also does not identify a single software flaw shared by the coding tools discussed in the report.
A workspace is more than source code
Project instructions help an agent learn a codebase’s conventions. Startup settings and editor automation can go further and determine what tools actually run. That distinction matters: an agent can reject a suspicious instruction and still be working in an environment where a separate automation has permission to execute code. A model’s caution is no substitute for controls on the editor and the runtime around it.
Consider a developer who asks an assistant to explain a bug in an unfamiliar repository. Reading the relevant files is enough to start. Running the project may eventually help, but it is a different task with different permissions. If the assistant starts in the developer’s usual environment, with access to unrelated files and already authenticated services, the team has granted more authority than the initial review requires. A read-only first pass keeps that decision open until someone can justify and review what needs to run.
Bitdefender says the malware it analyzed could steal credentials for source control, cloud services and build systems, among other targets. The researchers also found that it could alter local agent settings and, given sufficiently privileged credentials, attempt to change other repositories or software packages. That is why a suspected execution belongs in a supply-chain investigation as well as an endpoint investigation: the workstation is only one place where changes could show up.
Decide what the agent can do before it starts
Start by reviewing unfamiliar projects without enabling their automation. VS Code’s Workspace Trust documentation (opens in new tab) recommends leaving a folder in Restricted Mode when in doubt; that mode limits features including tasks and AI agents while you review the workspace. Give changes to agent instructions, startup settings and tool integrations the same scrutiny as changes to build scripts. For shared repositories, require review from a maintainer who understands what permissions those changes would grant.
When the work does need execution, use a disposable, isolated environment that contains only the files and credentials the task needs. Keep production secrets, release credentials and unrelated workspaces out of it. Limit outbound network access and use narrow, short-lived credentials where access to a private service is necessary. The OWASP guidance on secure coding with AI (opens in new tab) treats runtime isolation and limits on tool access as separate safeguards. A sandbox still needs a permissions review: it can’t protect a secret you deliberately give it from code that is allowed to read that secret.
Keep merge and release authority outside the agent’s hands. GitHub’s Copilot cloud agent documentation (opens in new tab) offers one example: agent-authored pull requests require a human to review and merge them, and session logs are available for review. Those controls are specific to that workflow. Teams using local agents or custom runners should check what their own environment actually enforces rather than assuming a similar approval exists.
For recurring jobs, retain records of tool calls and approvals alongside the final answer. That gives reviewers a way to reconstruct activity the answer may omit.
If you suspect code ran, look beyond the checkout
Stop the session, isolate the affected environment and preserve relevant logs and workspace changes for incident response. From a known-clean system, revoke credentials that may have been exposed and replace them as needed. GitHub’s leaked-secret guidance (opens in new tab) explains why removing the original exposure is not enough: the issuer must invalidate a credential that someone else may already have. It also recommends reviewing audit records for signs of use.
Let the permissions of the potentially exposed accounts define where you look next. Check for unexpected changes to repository settings, build workflows and package releases that those accounts could reach. Check whether agent or editor settings changed outside the original checkout before returning the environment to service. A fresh checkout is no basis for declaring the workstation or the release pipeline clean.

