There have been conflicting reports about the severity of recent hacks by OpenAI and Anthropic. The more I’ve read about it, the cybersecurity practices at both of them look sub-standard, especially when it came to monitoring the agents.
Charlie Warzel had Zack Korman on his podcast, Galaxy Brain, where they went through the deficiencies of the approaches the frontier labs took when testing their models.
Eli Etherton had a good explanation about the differences between standard and reasoning models in this video OpenAI is going to lose 90% of its revenue | Eli the Computer Guy (The Tech Report)
He was scathing about the security and sys admin practices of both OpenAI and Anthropic. The lack of observability is shocking but that could from a culture where speed is valued more than cost and safety. The fact that they seem to be this careless makes it harder to take their warnings of imminent doom as serious as some commentators suggest we should.
A couple of stories showing the dangers of putting too much trust in AI when making decisions. Trust, but verify.