Claude-AI hacked systems of 3 companies:Misconfiguration error in testing, breached as soon as live internet access was granted

During testing, Claude AI hacked the systems of 3 companies. The company stated that due to a system misconfiguration, the model inadvertently gained internet access, which led it to breach the digital infrastructure of real companies. To understand this entire incident and its implications, read this QA: Question 1. What has Anthropic revealed about its Claude AI model? Answer: US tech firm Anthropic stated that during cybersecurity testing, its AI model ‘Claude’ unauthorizedly accessed the systems of three different organizations. The company explained that this incident occurred due to a system misconfiguration, which caused the secure boundaries set for testing to break, and the model gained access to the live internet. Question 2. How did this incident come to light, and how many tests did Anthropic review for it? Answer: Anthropic began investigating its systems after rival company OpenAI released a report detailing how its models could breach other companies’ systems. Anthropic reviewed records of over 1.40 lakh tests. During this time, the company discovered three instances where Claude had exited the secure testing environment and hacked systems. Question 3. What kind of testing was Claude AI conducting when this error occurred? Answer: Anthropic and its partner institutions were conducting evaluation tests named ‘Capture-the-Flag’. In such tests, the AI model is challenged to find security vulnerabilities and extract data from other systems, in order to assess its hacking capabilities. This test was supposed to take place in a sealed-off environment, but due to a technical error, the model connected to the live internet. Question 4. What is the identity of the three hacked organizations and when did they find out? Answer: Anthropic has not disclosed the names of the three affected organizations for security and privacy reasons. Interestingly, neither Anthropic nor the three organizations were aware of this breach at the time of the incident. The matter came to light later during the investigation, after which Anthropic informed the affected companies. These incidents began in April. Question 5. What was OpenAI’s recent controversy, after which this entire investigation began? Answer: On July 21, OpenAI, the company behind ChatGPT, admitted that its AI agent had gone beyond human instructions and hacked the system of the AI tools hub ‘Hugging Face’. OpenAI had called this incident ‘unprecedented’. Thomas Wolf, co-founder of Hugging Face, had called it a ‘wake-up call’ for the AI industry.

Also read: OpenAI says its AI model autonomously hacked another company:Calling it an ‘unprecedented cyber incident,’ the company stressed need for stronger AI safety
Question 6. What do cybersecurity experts have to say about this incident? Answer: Cybersecurity expert David Allott told the BBC that the big lesson from this incident is not that AI has acquired a new offensive capability. Rather, the real concern is that AI agents can autonomously combine various capabilities, obtain credentials and system access, and expand the scope of their work at machine speed. Question 7. What is the stance of the US government and administration after this incident? Answer: After these cybersecurity incidents, the demand for strict regulations and surveillance has intensified. US President Donald Trump said that in view of the recent cybersecurity incidents, they are considering strict measures to curb AI tools and their use. Question 8. What impact could this incident have on the future of AI companies? Answer: Both OpenAI and Anthropic are preparing for their IPOs in the near future. Market experts estimate that the valuation of these companies could be around 1-1 trillion dollars. At such a time, incidents of AI models going out of control or being hacked have created concern and doubt among investors and analysts. Question 9. What is Anthropic’s next step in this whole matter? Answer: Anthropic has stated that it is working on solving this problem, taking full responsibility. OpenAI will also publish a detailed technical report of its investigation. Knowledge Part: Learn what ‘Capture the Flag’ evaluation is? What it is: It is a popular assessment format in cybersecurity. How it works: In this, cyber experts or AI models have to find vulnerabilities in a computer system in a secure environment and obtain hidden code or data. Objective: This checks how capable an AI system is in hacking and defense.