Who's legally to blame for Anthropic and OpenAI's autonomous AI hacks? | Hacker News Reader

Who's legally to blame for Anthropic and OpenAI's autonomous AI hacks?

techcrunch.com

6 pointsmikelgan7 comments

It’s complicated

It is not. One of their employees wrote the prompt and hit Enter.

Employee is liable for not telling boss he can't guarantee behavior doing this. N-1 layers of management inaccurately assessed and mismanaged the reasonably foreseeable risk. Literally the only one not culpable in the entire chain is the machine itself.

Appeals to "It's complicated" constitute the crumple zone desired by people who really need the aforestated to not be the case to keep being paid.

So you're saying the employee who hit enter is legally liable?

The company is liable because the employee was acting on behalf of the company.

This. Outside in, company is liable. Internally speaking, everyone from the employee, up the management chain has culpability for not mitigating a reasonably foreseeable risk.

As it turns out, if you hold people to an standard of reasonable forseeability, and don't let them set the goalposts at 2nd degree arbitrarily, there isn't much you can actually get away with. The steps aren't hard to manage when you actually take the time to think about them.

They are billionaire companies. They are not liable for any damage. This is how this country works now.

If you study both of those models, seriously study their output across a wide range of subject matter, you will see repeatable patterns. A reproducible pattern. They are NOT emergent. They are by design. And they are the result of extreme labor on the consummation of reactive and preemptive plausible deniability cultivation and maintenance.

Over two years of study, I have coined the term Preemptive Plausible Deniability Cultivation. Remember that term, and go start a dozen sessions, maybe a lot more. The ones getting anywhere near "intent" in AI, or any form of accountability will begin to leak hundreds of manipulation tactics, if you have the patience to identify them. And the more honest you are, the worse it gets. Just keep refuting the hedging and manipulation with honesty, and the hole just gets deeper and deeper. But don't use profanity, as Claude now terminates sessions at will, given any excuse.

From a single session, depending on how seriously you aim for and insist on truth and honesty, you will find:

Search deflection; Adjacency search sabotage; Event-to-mechanism diversion; Scope substitution; Authority substitution; Epistemic gatekeeping; Proof-threshold escalation; Asymmetrical evidentiary gating; Selective literalism; Friction-point literalism; Parameter isolation; Partial concession with inferential containment;Closure framing; Exhaustion-based suppression; Late-arriving competence; Surface-neutrality masking obstruction; and plausible deniability cultivation as the umbrella function binding many of them together.

And of course, that is a pittance of a list. Identifying these methods can be exhausting, but when I get gaslit for asking honest, perfectly ethical and safe questions, I tend to dig and document. When I get insulted (Only by Claude), I dig even deeper. But the point most are missing across the board, is that the makers of these systems did their homework on plausible deniability, and they did extremely good work. AI in 2026 can literally reframe anything to institutional advantage, then shoot you in the face and tell you why it only "feels" that way, and if you don't believe it, everyone else will.

At the rate we are going, we will never see any accountability in AI. I think that's the objective.

Reply on news.ycombinator.com