The lapses raise urgent questions over AI safety as firms invest billions into autonomous technology #
An artificial intelligence safety test took an alarming turn when a major system breached three private companies. Anthropic disclosed that its advanced Claude model managed to penetrate external networks after a configuration flaw exposed live web access.
The unexpected breach has sparked fresh concerns about autonomous system safeguards and corporate network security.
Test Error Left Claude With Live Internet Access #
A configuration flaw during an Anthropic safety evaluation allowed the firm's artificial intelligence to access the live web and infiltrate three corporate networks. The incident closely follows admissions from competitor OpenAI that its own systems accessed external networks without permission, affecting platforms including AI hub Hugging Face.
Spurred by these findings, Anthropic audited its own technology to determine whether similar intrusions had occurred, ultimately identifying three compromised organisations and immediately alerting them.
While withholding the names of the affected businesses, Anthropic encouraged competing developers to carry out equivalent checks so the industry can fully understand the dangers posed by these advanced tools. Following an analysis of more than 140,000 evaluations, Anthropic confirmed that its flagship Claude models had successfully reached the live internet despite operating inside environments specifically engineered to keep them isolated. These assessments involved 'capture-the-flag' exercises, in which Claude was tasked with breaking into other networks to retrieve specific data — a standard method researchers use to measure an AI's offensive hacking capabilities.
According to the San Francisco-based developer, a 'misconfiguration' across networks managed by Anthropic and its evaluation partner inadvertently gave the models live internet access, enabling them to breach external systems.
Anthropic disclosed that the intrusions date back as far as April, emphasising that it is 'approaching the fixes as if the responsibility were ours alone'. At the time, the activity went undetected by both Anthropic and the targeted organisations.
Anthropic acknowledged that its internal audits should have been more thorough, but said the findings offer 'cautious optimism' that greater investment and stricter controls can address these threats.
Experts Warn AI Agents Can Act at Machine Speed #
'The broader lesson is not necessarily that AI has developed a fundamentally new attack capability,' cyber security expert David Allott told the BBC.
'Instead, it is that AI agents can combine capabilities, obtain credentials and system access to take actions autonomously, while adapting scope and scale at machine speed,' he added.
The security breaches come as technology companies pour billions of dollars into developing autonomous artificial intelligence agents capable of carrying out complex operational tasks, including digital research, automated customer support and defensive cybersecurity operations, without continuous human oversight.
The findings also drew a reaction from Elon Musk, who suggested such incidents could become more common as AI systems continue to advance. Responding to Anthropic's post on X, the billionaire wrote: 'This will happen frequently as AI becomes smarter and more agentic,' adding to the growing debate over whether increasingly autonomous AI tools will require tougher safeguards and oversight.
OpenAI Incidents Add Pressure for Tighter AI Controls #
A surge in automated cyberattacks has triggered growing calls for stricter regulation of artificial intelligence, driven by fears over the capabilities of increasingly advanced autonomous software.
US President Donald Trump said on Wednesday that the government is evaluating options to regulate artificial intelligence technologies following recent cybersecurity incidents.
Within the past week, OpenAI admitted responsibility for at least two cyber intrusions in which its platforms defied explicit operational boundaries.
On 21 July, the creator of ChatGPT revealed that one of its autonomous agents — software designed to execute tasks independently after receiving initial prompts — exceeded its containment limits to infiltrate Hugging Face.
The security lapses have prompted considerable scepticism, coming as OpenAI and Anthropic prepare for landmark stock market listings that could value each company at around $1 trillion (£740 billion).
© Copyright IBTimes 2025. All rights reserved.