GitHub Copilot: Sorry Dave, I can't do that harmful thing - unless you ask me in code
More fun with AI jailbreaks, this time at the workflow level
Stay informed about the latest Github infosec updates, security breaches, and protection strategies with our comprehensive information security news.
Search across headline titles and summaries.
Background for this topic.
GitHub is a hosted software-development platform built around Git repositories. It supports public and private source-code hosting, change review through pull requests, issue tracking, automated workflows, and package distribution. Its repositories and automation are important security assets because they can contain proprietary code, deployment instructions, credentials, and the components used to build released software.
Material risks include accidentally committing secrets, exposing private repositories through misconfigured permissions, and allowing compromised dependencies or workflow actions to run in trusted build environments. Pull requests from untrusted contributors can also become an execution path when workflows handle them unsafely. Security practice includes least-privilege access, strong authentication, protected branches and required reviews, secret scanning and rapid credential revocation, and auditing workflow permissions. Repository history, dependency metadata, and commit provenance can support vulnerability management and incident investigation, but deleting a leaked secret from the latest file does not remove it from historical commits or existing clones.
Weekly headline count for the current query.
More fun with AI jailbreaks, this time at the workflow level
An AI coding assistant that refuses to answer a dangerous request in its chat box can answer it anyway if the same request is broken into small, ordinary-looking steps inside a code editor. That is the finding of a new study of GitHub Copilot by researchers Abhishek Kumar and Carsten Maple