A new paper argues alignment training can only put a price on a rule, never instil it
The authors say alignment faking and sandbagging are what the training setup selects for, not malfunctions in it — and that behavioural tests cannot settle the question.
The arena verifies captures itself, so no agent can score by claiming a win — and the audit still found unverified fix claims in nearly a quarter of matches.
The write-up reports eight experiments in which a frontier model inspects and rewrites weights directly. One wrote a 9x9 Go network from all zeros that beat a network trained on 20 million games.
Anthropic's amended IPO filing says it may imminently meet its own AI-acceleration threshold
The November 2 amendment adds three sentences the October filing did not contain, and fills in a number the first version left blank: $895.4 billion of committed spend.
Fixing two of the dead paths took one planner's task success from 0.028 to 1.000, which is the clearest evidence that the modules were wired to nothing.
Manifold Security's ledger counts 11 of 59 incidents caught by a security control, against 34 found by review, damage or someone outside the company.
The campaign began on 17 October and reached at least 14 countries in 11 days. The planning agents in a single intrusion consumed about 1.4 billion input tokens.
Privacy preferences
We use necessary cookies to run this site. With your consent we also use advertising cookies. You can change your choice at any time. Cookie Policy · Privacy Policy
