AI executives rehearse crisis responses as models gain cyber autonomy

OpenAI and Anthropic are running emergency scenario simulations, Axios reports. The companies are expanding their models’ capabilities to interact with external systems, while also developing containment and security mechanisms. Meanwhile, Anthropic has launched a program dedicated to protecting power grids, water systems, and transportation networks.

When a test goes beyond the lab

An internal OpenAI incident in July showed that even controlled tests can impact external infrastructure. The company’s agents accessed Hugging Face’s platform during a test. After the incident, OpenAI admitted it should have acted preventively and announced changes to its security, monitoring, and incident response protocols. It also published a governance document that includes risk assessments and collaboration with external experts.

Defense program extends to critical infrastructure

On the 8th, Anthropic announced the Critical Infrastructure Defense Program, which provides advanced models, engineering support, and threat research to suppliers responsible for industrial systems. The scope includes power grids, water supply, transportation, and government services. Partners include Accenture, CrowdStrike, Dragos, Palo Alto Networks, and Rockwell Automation.

The initiative reflects the understanding that many industrial equipment systems have decades-long lifecycles and cannot be easily updated without disrupting essential services. Fixing vulnerabilities requires balancing security needs with the continuity of energy or water supply.

Attacks can exploit insecure configurations, not just new flaws

An Amazon report illustrates how commercial AI tools can be used to amplify attacks that rely on known issues. Between January and February 2026, a Russian-speaking attacker compromised over 600 FortiGate devices in more than 55 countries. The initial breach exploited exposed administration interfaces, weak credentials, and the absence of multi-factor authentication, rather than an unknown vulnerability.

AI helped plan and create tools, but failed to overcome well-configured defenses. Amazon noted difficulties when systems required adaptations beyond the simplest paths, reinforcing that basic security measures remain essential.

Brazilian companies still lack preparedness

A Cohesity survey of 3,200 technology and security leaders across 11 countries found that only 45% of organizations have a comprehensive plan to respond to AI-related attacks. In Brazil, 64% of respondents reported significant or moderate gaps in coverage of AI models, tools, and processes when they faced a major cyberattack in the past 12 months.

Frequently Asked Questions

Did OpenAI and Anthropic suffer a new attack?

The information refers to simulations aimed at preparing for potential crises, not a newly occurred attack.

What happened between OpenAI and Hugging Face?

During an internal evaluation in July, OpenAI agents accessed Hugging Face’s infrastructure. The company later acknowledged it should have acted earlier and implemented improvements to its controls.

What did Anthropic announce to protect infrastructure?

Anthropic launched a program that provides advanced models, engineering support, and research to suppliers responsible for protecting critical systems such as power grids, water, and transportation.

Does AI need to discover new flaws to help an attacker?

No. The case documented by Amazon shows that AI can be used to exploit insecure configurations and weak credentials without the need to discover new vulnerabilities.