GitGuardian
CompanySecurity timeline
- Malware reported by Help Net SecurityWeek in review: secret stealer hits GitHub repositories and Claude Code turned into a remote-controlled implant
- Vulnerability reported by CSO OnlineClaude Code permission files leak thousands of secrets from public repositories and Docker images
Recent news
Week in review: secret stealer hits GitHub repositories and Claude Code turned into a remote-controlled implant
A Help Net Security weekly review reports a secret-stealing threat affecting GitHub repositories and says Claude Code was turned into a remote-controlled implant. It also features an interview with GitGuardian CEO Eric Fourrier on leaked credentials, personal liability for security leaders and how to brief a CEO about exposed keys. The teaser gives no technical details, so the full articles are needed for specifics.
Attackers abuse open-source AntiFox AI pentesting tool to reach AWS production secrets
Arctic Wolf linked AntiFox, an open-source AI penetration-testing platform, to attacks on internet-facing applications and their AWS environments, with GitGuardian also reporting the activity. Attackers deployed an ASPX webshell on a Microsoft Internet Information Services server, pulled database connection strings from configuration files, and then used valid IAM and EC2 role identities to probe S3, Secrets Manager and other services. AWS said its services were not compromised and that the attackers used credentials taken from customers' applications.
AI hacking frameworks found running on internet-facing servers, some without authentication
GitGuardian reported that AI penetration-testing frameworks, including PentAGI, ARTEX and Agent Zero, are exposed on internet-facing servers, with some consoles open without authentication. A November 2026 Shodan snapshot recorded 1,428 ARTEX-family, 419 PentAGI and 173 Agent Zero services, though the counts vary by query and should not be combined. The researchers stress that these are dual-use tools and an access-control problem rather than a new malware family.
Claude Code permission files leak thousands of secrets from public repositories and Docker images
GitGuardian found 31,088 distinct secrets in .claude/settings.local.json files across 23,276 public GitHub repositories, with 5,343 still valid when detected. When a developer picks "Yes, and don't ask again", Claude Code stores the full command string, including any embedded credentials, as an allow rule in that file. A further 613 secrets were found in public Docker Hub images, 99 of them valid, because COPY instructions pulled the file in and Docker does not respect .gitignore. Anthropic's Claude Code lead said the company is exploring built-in secret scanning and an auto mode.
GitGuardian CEO on leaked credentials, personal liability and measuring security progress
In a Help Net Security interview, GitGuardian CEO Eric Fourrier discusses leaked credentials and personal liability for security leaders and boards. He covers how fear of blame pushes mistakes out of sight, how to brief a CEO on exposed keys, and how to measure progress beyond alert counts. The teaser also touches on disclosure deadlines and new risks from AI tools and agents.