Attacking AI Agents: The Best Defense Is Still Basic Security

The first documented attacks using artificial intelligence agents leave a lesson that isn’t flashy but is very practical for systems and cloud teams: agents don’t open new holes, they find the existing ones faster. That’s the takeaway from the BOSSAI case, described by Acronis in November 2026, and from the first agent-driven attack reported to Spain’s Data Protection Agency (AEPD), which consultancy Logixs covered in an analysis and which OpenSecurity broke down in detail.

AI agent cyberattacks in 30 seconds

  • Acronis documented an attacker using 12 AI agents against hundreds of organizations in 11 countries, with about 10 confirmed intrusions.
  • The breaches came through exposed secrets and poor API access control.
  • Where there was strong authentication, the agents didn’t get in.
  • Shadow AI and AI-generated code are widening the attack surface from the inside.

For anyone running infrastructure, the useful question isn’t whether AI will make attacks more sophisticated, but which part of the environment would be exposed if someone could scan all of it, in parallel, without ever getting tired.

Where the agents got in

According to Acronis’s Threat Research Unit, the operator behind BOSSAI, based in Taiwan, split the work across nine active agents: some maintained the attack infrastructure, others scanned targets, and others tried to exploit flaws. Over 12 days of observation it targeted emergency services, banks, universities, and military systems, and the roughly 10 confirmed intrusions in Taiwan, Mexico, and Thailand included emergency records from seven Taipei districts.

The attack vector was nothing exotic: poor API access control and exposed secrets. The agents also had clear limits, since they didn’t coordinate on their own (a human passed results from one to the next) and they went as far as hallucinating vulnerabilities that didn’t exist. In the Spanish case, the AEPD reported in September that an agent built on a well-known language model found flaws, accessed the system, and pulled personal data and invoices from an unidentified organization.

What to check in your infrastructure

The measures that stopped BOSSAI are the usual ones, applied rigorously:

  • Secrets out of code and repositories. Secrets managers, key rotation, and repository scanning for leaked credentials.
  • APIs with real authentication and authorization. An inventory of exposed endpoints, rate limits, and per-object permission reviews.
  • MFA and least privilege on management panels, cloud consoles, and service accounts.
  • Your own agents under control. Every internal agent with its own identity, minimal permissions, action logging, and testing in an isolated environment before touching production.

The credentials problem is very real in the cloud: we’ve already covered how a stolen key can turn the use of AI models into a million-dollar bill through LLMjacking.

Shadow AI: the risk that’s already inside

Logixs’s analysis also focuses on the internal front. It cites a Cegid study finding that 8 out of 10 Spanish employees who use AI at work turn to tools their company hasn’t provided, and Veracode’s 2025 GenAI Code Security report, which found security flaws in 45% of code samples generated by more than 100 models. Logixs also brings figures from its own audits, with no published methodology, which should be treated as an indicator rather than a statistic.

Its proposed response is to offer approved AI tools with clear rules, so employees don’t have to look for alternatives on their own, and to continuously review AI-generated code, with detection of keys and personal data from the very first line. For the financial sector and essential operators, DORA and NIS2 already require managing these risks.

Frequently asked questions

How did BOSSAI’s AI agents get in?

According to Acronis, by exploiting poor API access control and exposed secrets. Where there was strong authentication, they failed to get in.

How many organizations did BOSSAI attack?

Hundreds of organizations in at least 11 countries, with about 10 confirmed intrusions in Taiwan, Mexico, and Thailand.

What should a systems team check first?

Exposed secrets in code and repositories, API authentication and permissions, MFA on consoles and service accounts, and the permissions of their own AI agents.

What is Shadow AI?

Employees’ use of AI tools that their company hasn’t approved or doesn’t control.

Sources:

This article is also available in Spanish: “Agentes de IA atacantes: la defensa sigue siendo la seguridad de siempre,” on revistacloud.com.

Scroll to Top