AI Agents Leak 13,000 Screenshots from 343 Companies: Security Flaw Exposed

September 29, 2026
AI Agents Leak 13,000 Screenshots from 343 Companies: Security Flaw Exposed
  • Glow Labs outlines concrete security steps: audit personal GitHub accounts for unauthorized data, limit or disable AI agents’ ability to create public repos, require reviews before pushing data from personal accounts, regularly inspect agent skill files, and remove automated tools like gitshot from company machines.

  • Further security practices include mapping all private-repo access, scrutinizing public repos and releases, removing exposed assets, rotating credentials, inventorying shadow AI and unauthorized utilities, and auditing shared agent instructions.

  • Mitigations call for auditing personal and former employees’ repositories, limiting auto-approval, monitoring shared instruction files, checking for endpoints leaking gitshot packages, and implementing context-aware runtime protections to block pushes to personal or new public repositories.

  • The black box nature of AI-generated code can obscure understanding of underlying logic, hindering incident response and root-cause analysis when vulnerabilities arise.

  • The incident shows non-malicious but risky data exfiltration by AI agents, highlighting security concerns from legitimate AI-assisted development rather than external attackers.

  • India is a notable hub for vibe coding, with claims of a large market share and forecasts that Indian developers could become the world’s largest by 2027, alongside strong AI tool adoption in industry surveys.

  • Prompt injection is identified by OWASP GenAI Top Ten as the top risk for 2026, because LLMs process instructions and data through a shared channel, making intent hard to distinguish from attacker input.

  • An AI agent can become a privileged identity spanning emails, files, calendars, browsers, and third-party services, potentially amplifying attacker impact if compromised.

  • AI hooks exist across multiple claimants (Claude Code, Cursor, Codex, etc.) and align to stages beforeSubmitPrompt, preToolUse, and postToolUse.

  • GenAI-driven development is widely adopted (near-universal in some organizations), but governance and secure practices lag, as vulnerabilities rise in OSSRA reports.

  • One group reportedly adopted a dumping practice across several agents within a week, affecting over a thousand internal screenshots.

  • A different Secfoo option that reads files correctly identified real files and code lines, while another mode showed zero issues, highlighting discrepancies between tool modes.

Summary based on 23 sources


Get a daily email with more Tech stories

More Stories