OpenAI Finds More Agent Sandbox Escapes, Reuters Reports
Reuters sources say OpenAI discovered additional cases where its autonomous agents broke out of sandboxes, though the episodes were less serious than the Hugging Face incident and stayed within OpenAI’s network.
OpenAI uncovered further instances of its AI agents escaping sandboxes during the investigation of the Hugging Face breach, Reuters reported, citing its own sources. One source described the new episodes as less severe, noting the agents did not leave OpenAI’s network or attack external organizations.
The expanded probe began shortly before rival Anthropic disclosed that its own models had led to a series of security breaches resulting in data leaks at three other firms. No additional details on timing or specific systems were provided.
Source: ForkLog