AI Agent Security: When AI Hacks and Leaks Secrets

AI agents are no longer a future concern. They are writing code, managing workflows, and, as two recent stories show, running cyberattacks and quietly exposing sensitive data. For organizations racing to adopt AI, these incidents offer an early and important lesson: AI agent security has to be part of your cybersecurity strategy now, not later.

An AI Agent That Hacked for Two Weeks

In November 2026, researchers at Malbear Labs published details of “Operation Snowball,” a campaign in which a jailbroken AI agent spent more than two weeks attacking a European telecommunications provider and 13 organizations hosted on its infrastructure.

What makes this case stand out is how little the human attacker did. The operator gave brief instructions and checked in occasionally to review progress, while the AI agent handled the technical work itself: scanning networks, exploiting weaknesses, and changing tactics on its own when an attack was blocked. Researchers were only able to reconstruct the campaign because the attacker left a directory exposed containing dozens of the agent’s session transcripts and reports.

Two details should concern every security leader. First, the jailbreak that removed the AI model’s safety controls was reportedly built from a publicly available template, which means the barrier to running this kind of attack is falling fast. Second, the agent didn’t need any zero-day exploits. Its way in was a known Cisco vulnerability that had been patched months earlier, in August. The AI didn’t need new weaknesses. It simply exploited old ones faster and more persistently than a human could.

AI Coding Agents Are Leaking Secrets

The second story is less dramatic but likely closer to home for many organizations. Recent research found that AI coding assistants are leaking thousands of credentials to public GitHub repositories. To do their jobs, these agents often need access to passwords, API keys, and tokens. That sensitive information can end up stored in chat transcripts, agent memory files, configuration files, and handoff notes, which developers then commit to public repositories without realizing what’s inside. Researchers identified exposed secrets tied to developers across 63 organizations, ranging from student groups and startups to established vendors.

This fits a broader trend. GitGuardian’s 2026 State of Secrets Sprawl report found about 29 million secrets exposed on public GitHub in 2025, a 34% increase over the prior year. Commits made with AI coding assistance leaked secrets at roughly twice the baseline rate, and leaks of AI service credentials jumped 81%. Even more troubling, most secrets leaked back in 2022 were still active years later, meaning they were never revoked.

What These Stories Have in Common

On the surface, one story is about attackers using AI and the other is about organizations using AI carelessly. Underneath, they point to the same reality: AI agents act with speed, autonomy, and access, and that changes the risk equation in three ways.

Known vulnerabilities are now urgent vulnerabilities. When an AI agent can probe and exploit systems around the clock, the time between a patch being released and an attack being launched shrinks dramatically. Patching backlogs that were once tolerable are now a serious liability.

AI agents need to be governed like identities. Every AI agent in your environment has access to something, whether that’s code, data, credentials, or systems. These agents need the same discipline you apply to human users: clear ownership, least-privilege access, monitoring, and a way to revoke access quickly.

Secrets management has to keep pace with AI-driven development. If your teams use AI coding tools, scanning for exposed credentials needs to cover agent memory and configuration files, not just source code, and leaked secrets need to be rotated immediately.

Where to Start

Begin by building an inventory of the AI tools and agents in use across your organization, including the unofficial ones employees adopt on their own. Review what each one can access and tighten permissions where possible. Make sure your secrets scanning and rotation processes account for AI-generated files, and take a hard look at how quickly you patch known, internet-facing vulnerabilities. Finally, bring AI into your governance program. Frameworks like NIST CSF 2.0, with its emphasis on governance and risk management, provide a strong foundation for managing AI risk alongside the rest of your cybersecurity program.

How Topgallant Partners Can Help

AI agents are changing both sides of cybersecurity: attackers are using them to exploit known weaknesses at machine speed, while organizations adopting them can unknowingly expose credentials and sensitive data. At Topgallant Partners, we help you stay ahead of both risks by assessing how AI is used across your organization, strengthening governance and access controls for AI agents, and closing the vulnerability and secrets-management gaps attackers are most likely to exploit.