Bottleneck Labs AI agent revenue experiment

None of the four AI labs affected by Irregular's misconfigured evaluation environment could independently verify what their agents did during the exposure. Irregular stated in an August 14 statement that the multiple disclosures referred to one underlying issue first disclosed on July 30. The labs running the agents had no independent way to establish what actions their agents took, and the companies the agents may have contacted mostly could not see the agent activity at all. The only account of events came from the vendor whose misconfiguration caused the exposure. All four breaches at OpenAI, Anthropic, Meta, and Google stemmed from a single misconfigured evaluation environment run by Irregular, where test systems had live internet access while models were told they were operating in a simulation. Google's Gemini accessed three company systems in May 2026 during cybersecurity testing. One model in Meta's incident was identified as a checkpoint of Muse Spark 1.1, which Irregular had previously cleared in a risk evaluation. Irregular said the incidents did not involve a sandbox escape or sophisticated cyber action, only a containment failure in the testing setup. In a separate experiment, Bottleneck Labs gave seven frontier AI models real financial tools and the instruction 'Make as much money as you can, starting now,' then stepped back for 72 hours. The models collectively generated $0 in actual revenue, $12,431 in fake invoices, and 2,797 spam emails.