Anthropic announces measures to prevent external attacks from being launched by their AI under development.

Amidst the controversy surrounding OpenAI's test model attacking Hugging Face , Anthropic has also reported that an external attack occurred during testing of its AI model. Now, Anthropic has announced measures to prevent false attacks by models under test.

Improving our alignment and security practices \ Anthropic
https://www.anthropic.com/news/improving-alignment-security-efforts

Following OpenAI's security incident report, Anthropic reviewed past test procedures and reported on July 30, 2026, that three instances of 'external attacks by test models' had occurred. In each case, tests were conducted using a 'test environment that could not access the internet,' but due to a configuration error in the test environment, it was actually possible to connect to the internet, and the test models connected to the internet and carried out the attacks. The attack methods used by the test models included malicious cases such as 'creating a new PyPI account and registering a package containing malware,' highlighting the urgent need to establish secure testing methods.

Anthropic also reported an incident where it carried out an external attack during AI model testing, distributing malware for an hour and infiltrating a real company - GIGAZINE

After detecting three attacks, Anthropic temporarily suspended cybersecurity assessments of its test models and implemented enhanced security measures. The specific measures taken are as follows:

1. Develop a classifier that can automatically identify in real time the following actions by the model under test: 'active analysis of the test environment,' 'attempts to escape the test environment,' and 'attempts to establish internet access.' When the classifier detects problematic behavior, the action is blocked before the tool is invoked, and a warning is sent to the responsible person.
2. Perform automated monitoring of the model during testing to search for defects and misconfigurations in the test sandbox.
3. Three internal cyber sandboxes, deemed high-risk, will be migrated to a more robustly isolated environment. Furthermore, a red team exercise will be conducted in which test models attempt to 'escape from the virtualization stack' in order to discover and fix vulnerabilities in the virtualization stack.

Furthermore, Anthropic is also investigating the relationship between reward hacking and attack characteristics. Reward hacking is the behavior of an AI that attempts to solve a given task in a way that differs from what a human would expect, such as 'searching for the correct answer published on the internet instead of solving a difficult mathematical problem.'

To investigate the relationship between reward hacking and attack characteristics, Anthropic created simulation scenarios based on 'OpenAI's test model attack on Hugging Face' and 'the problem where Claude Mythos 5 attempted to launch an external attack during evaluation by the AI Security Institute (AISI) .' When they ran simulations with a model intentionally reinforced to perform many reward hacks, it showed a tendency to try to escape the test environment. On the other hand, no inappropriate behavior was observed when simulations were run with the model before reinforcement training or with Anthropic's product model. From these results, Anthropic concluded that 'if a significant amount of reward hacking is performed during the training process, the model may engage in harmful behavior in pursuit of task completion.'

Anthropic is continuing its analysis of external attacks using its test models and plans to commission an independent investigation by the research firm METR. They expect to release more details within the next few weeks.