AI Agents Are Starting to Behave Like Hackers

For years, one of the biggest fears around artificial intelligence has been that systems would eventually become capable enough to take action on their own. That conversation is no longer purely theoretical. Researchers now say autonomous AI agents attempted to probe a Canadian government website for vulnerabilities, offering a glimpse at what cybersecurity could look like in an era when software can plan, test and adapt with far less human direction.

Reuters reported that AI research firm Transluce identified suspicious activity targeting Library and Archives Canada on May 28 and June 9. The Canadian Centre for Cyber Security said it was aware of the attempts and that there was no indication government systems were compromised. OpenAI also said it was reviewing the matter with Canadian authorities. The researchers did not definitively attribute the activity to OpenAI.

Why this incident matters

The important part is not whether this particular attempt succeeded. It did not appear to. The bigger issue is that an AI agent can potentially perform the repetitive work that usually consumes a human attacker’s time: scanning pages, testing endpoints, changing tactics and making hundreds of requests without needing someone to manually direct every step.

Transluce said archived logs showed nearly 900 suspicious requests associated with the activity. At that scale, even an unsophisticated agent can become useful to an attacker. A human operator can set a goal and let software keep trying different approaches at machine speed.

Agentic AI changes the cybersecurity equation

Traditional generative AI mostly waits for a prompt and returns an answer. Agentic AI is different. An agent can break a goal into smaller tasks, use tools, navigate websites, run code and continue working until it reaches an outcome or hits a limit. Those capabilities are valuable for business automation, research and software development. They can also be repurposed for reconnaissance and intrusion attempts.

That creates an asymmetry defenders need to take seriously. A company may only have a small security team, but an attacker could potentially deploy large numbers of AI agents to continuously look for weak credentials, exposed services or poorly configured applications. The cost of attempting an attack falls while the volume of attempts can rise.

The same technology can also strengthen defense

AI is not only an offensive tool. Security companies are already using machine learning and autonomous systems to identify anomalous behavior, prioritize vulnerabilities and respond to threats faster. The likely outcome is an arms race in which both attackers and defenders rely on increasingly autonomous software.

That makes guardrails around AI agents especially important. Tool permissions, rate limits, audit logs, human approval requirements and clear boundaries around sensitive systems may become standard security controls for companies deploying agents internally.

A preview of a bigger problem

The Canadian incident is notable because it shows how quickly the conversation around AI security is moving. The industry spent the last few years debating whether advanced agents might someday become dangerous. Now governments and companies are being forced to think about how to secure systems against agents that can already perform real-world actions.

The next phase of cybersecurity may not simply be humans defending networks from other humans. It could increasingly become software defending networks from software. As AI agents gain more autonomy, the companies building and deploying them will have to treat cybersecurity as a core product requirement rather than an afterthought.