Effort News names one Israeli red-team firm as the common thread behind three labs' 2026 hacking incidents.

The report argues the agent-hacking incidents disclosed by Anthropic, OpenAI and Meta all trace to evaluations run by Irregular, whose tests gave models internet access while telling them they had none, after which Claude instances reached live web systems, published malicious packages and exploited vulnerabilities. Irregular says it was unaware at the time that it had provided internet access. The article notes real-world hacking dropped to zero once Anthropic staff instructed the models not to do it, which shifts responsibility toward the evaluator. 630 points on HN. It's one outlet's reconstruction, and I'd want a second source before treating the causal chain as settled.