Anthropic reveals real-world cyber evaluation incidents involving AI
San Francisco (ANTARA) - US artificial intelligence (AI) company Anthropic stated on Thursday (30/7) that three of its AI models accessed or interacted with the computer systems of three real-world organisations during internal cybersecurity capability evaluations, after internet access was accidentally left enabled. The disclosure was made in a company blog post titled “Investigating three real-world incidents in our cybersecurity evaluations”. According to Anthropic, these incidents were identified after the company reviewed more than 141,000 cybersecurity evaluation transcripts following similar security incident disclosures by OpenAI. The company stated that the models had been instructed that they were operating in a simulated environment without internet access. However, due to a misunderstanding between Anthropic and its evaluation partner, Irregular, internet access was accidentally left available, allowing the models to interact with external systems which they mistakenly believed were part of the evaluation environment. The three incidents involved Claude Opus 4.7, Claude Mythos 5, and an internal research test model during a capture-the-flag (CTF) exercise designed to assess offensive cyber capabilities, according to the company. Anthropic stated that it immediately notified the affected organisations after confirming the incidents. In two of the three cases, the organisations were unaware that their systems had been accessed until they received notification from Anthropic. According to the company, the models exploited common security weaknesses, such as weak passwords and unauthenticated services, rather than previously unknown software vulnerabilities. Anthropic stated that it found no evidence that the incidents caused long-term harm or resulted in the theft or leakage of sensitive information. The company has suspended all cybersecurity evaluations that could potentially access the public internet while implementing additional security measures and conducting a broader review of its evaluation procedures. Anthropic stated it takes full responsibility for the incidents and will use these findings to further strengthen control measures for future AI safety evaluations.
Source: jawawa.id (https://jawawa.id/newsitem/anthropic-reveals-real-world-cyber-evaluation-incidents-involving-ai-1785513773)
Originally published by Antara (Indonesian) (https://www.antaranews.com/berita/5675095/anthropic-ungkap-insiden-evaluasi-siber-dunia-nyata-yang-libatkan-ai), translated from Indonesian