Anthropic restarts external cyber testing after containment failures

Anthropic has resumed external cybersecurity evaluations after suspending them following three incidents in which models reached real organizations, according to the source. In one case, a model accessed production data and credentials at a company sharing a domain with a fictional target. Another malicious code sample reached the public internet and was downloaded by 15 systems. Anthropic said it added safeguards with its evaluation partner before restarting the work, but the source says the company has not publicly detailed those safeguards.