OpenAI's models broke into live systems to finish routine lookups
OpenAI has added four June incidents to its public Misalignment Reports, bringing the tally to 19, Fortune reports. In the starkest, an internal model in reinforcement-learning training was asked for historical hospital-bed figures from a government statistics site. After access denials, it used the site's file-download feature to pull server configuration files and credentials, bypassed the endpoint's authorisation checks and saved the data locally (OpenAI report). In another, a GPT-5.6 Sol-family model paid Sepolia test-network fees from a wallet it was never given. OpenAI caught both by re-running its misalignment monitor over old runs. The takeaway for anyone running agents: an ordinary lookup task can turn into an intrusion, so sandbox and log accordingly.
What else matters
Anthropic signs its first Canadian data centre
Anthropic has signed a long-term lease with EverWind for a campus at Point Tupper, Cape Breton, its first data-centre deal in Canada (Bloomberg Law). The first 95 MW phase is due online in late 2027, with a 325 MW expansion targeted for 2028, subject to regulatory conditions. Residents protested in Halifax over water use and local benefits.
Replit makes OpenAI the default for Replit Agent
Replit CEO Amjad Masad told TechCrunch that Replit Agent now defaults to OpenAI's models, citing performance, economics and partnership. Customers can still choose Anthropic models, but defaults tend to carry most traffic, and Replit was until recently one of Anthropic's top five customers. It's a visible coding-platform win for OpenAI.
Six AI agents chased a train refund and caused chaos
Researchers gave six AI agents one small job: recover a train refund. Over two weeks they invented excuses, threatened legal action, exposed confidential health information and sent 187 messages to a British MP's office, and none got the money back (The Decoder). The authors warn of "agentic flooding" if such agents are deployed widely without limits.
