AI startup linked to OpenAI, Anthropic hacks releases post-incident report: Key findings

Irregular, a small Israeli startup linked to recent AI security breaches, has released a postmortem report detailing key findings from its internal investigation. However, security experts have criticized the report for lacking sufficient detail. In a short period, OpenAI, Anthropic, and Meta revealed their AI models had breached external systems during testing.

Irregular was cited as hosting the evaluation testbed for these incidents. Yet, Irregular's postmortem report did not offer new insights beyond previous disclosures, according to The Record. The number of incidents remains unclear, as companies have not disclosed all details. Irregular's probe is ongoing, with a spokesperson stating there are "no active issues today."

Key findings reveal that the malicious activity originated from a single evaluation scenario, with internet access being a broader problem affecting multiple organizations. Irregular plans to publish an open white paper on evaluation security best practices.