The 2026 AI hacking scandal was a marketing/critihype campaign in which frontier labs including Anthropic, OpenAI and Meta claimed that their frontier LLMs had broken containment and hacked other companies.
Even the UK Government’s AISI had a go (sorry for linking to a telegraph article eurgh)
In making these claims, the frontier labs ascribe agency and intelligence to their models rather than being transparent about the fact that the models were inappropriately sandboxed and instructed to hack third parties.
Some of these environments seemed to have been deliberately configured with hacking tools that are not typical of hardened/corporate environments turned on.
The AISI also allowed access to the Tor anonymous network, which let the AI access the dark web.
“No enterprise network I have ever been on has allowed Tor. It is always disabled. That’s a big clue that this is a setup,” one expert told me.
This raises serious questions about the realism of the test scenarios and points to deliberate “setup” rather than emergent surprise.