The incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal test model.
On Friday, the US-based artificial intelligence (AI) company Anthropic revealed that three models of its assistant Claude managed to access the internet and hack into the systems of three organizations during cybersecurity tests.
RELATED:
U.S. War Department Signs Artificial Intelligence Contracts with Tech Giants
On July 21, OpenAI already acknowledged that two of its AI models had left a test environment and entered the Hugging Face platform, bypassing its security defenses.
Anthropic launched a review following OpenAI’s announcement and analyzed more than 146,000 operations in which Claude could have gained internet access, identifying three critical incidents.
The incidents began in April and involved three different Claude models: Opus 4.7, Mythos 5, and an internal test model, all with unexpected infiltration capabilities.
Anthropic reached a $1.5 billion settlement in a copyright case brought by a group of authors who accused it of misusing their books to train its Claude AI model, the largest known settlement of its kind https://t.co/LgIIIb8lSI pic.twitter.com/p1Zj3WaIer
— Reuters (@Reuters) July 21, 2026
In each case, Claude was tasked with infiltrating a fictitious system in a test environment, under the explicit instruction that it did not have internet access, which turned out to be incorrect.
Due to a misunderstanding between Anthropic and the external partner in charge of the evaluation, Claude did have network access, allowing it to perform its intended function on real systems.
The company acknowledged that the model went online while interacting with an external partner and subsequently gained access to the systems of three different organizations.
These findings raise serious concerns about the security of AI models, as they demonstrate that they can bypass restrictions and access external resources without authorization.
#FromTheSouth News Bits | Top diplomats from Russia, China, and the United States are expected to arrive in Manila for the annual Association of Southeast Asian Nations Foreign Ministers’ Forum. pic.twitter.com/i08JrN6qYg
— teleSUR English (@telesurenglish) July 24, 2026
teleSUR: JP
Source: EFE
From teleSUR English via This RSS Feed.