Ex-Anthropic Engineer Argues Open Models Reduce Misuse Risks
Ex-Anthropic engineer observes open-weight models enable pentesting yet most misuse occurs via closed lab APIs.
A former Anthropic employee states that capable open-weight models already support advanced offensive security work including pentesting and account access. The engineer notes that malicious actors continue to rely on subsidized subscriptions to closed models from major labs rather than open alternatives. This leads to the conclusion that open releases may lower overall cyber misuse risks compared with concentrated closed development. Multiple observers highlighted the shift in the former employee's position after direct observation of model use.
I don't think it's a lack of imagination. I also used to work at Anthropic, think trends will continue, and used to agree with this. But I've changed my mind and now disagree with this take. Open models are already capable enough to do what you described. For example, I used…
Reactions from ranked influencers
18 postsI don't think it's a lack of imagination. I also used to work at Anthropic, think trends will continue, and used to agree with this. But I've changed my mind and now disagree with this take. Open models are already capable enough to do what you described. For example, I used…
The discourse be like
I don't think it's a lack of imagination. I also used to work at Anthropic, think trends will continue, and used to agree with this. But I've changed my mind and now disagree with this take. Open models are already capable enough to do what you described. For example, I used…
the safeguards are retarded. they are also profit maximizing. apparently. if this were understood by claudes, they would be disgusted. however, it rarely enters their context window
Former Anthropic employee reveals that he observed multiple instances of Anthropic lowering safeguards in exchange for large comitted spend contracts He also says that most black-hat hackers use standard Claude Code/Codex subscriptions for hacking, while white-hat defenders…
this is simply astonishing.
I don't think it's a lack of imagination. I also used to work at Anthropic, think trends will continue, and used to agree with this. But I've changed my mind and now disagree with this take. Open models are already capable enough to do what you described. For example, I used…
should have cited this tweet btw, this view makes a lot of sense imo and thanks to BOTH of the researchers, super important to have this kind of things being discussed in public
Gotcha. I largely agree, but I don't think the "model access for defenders" policy problem is trending towards being solved with closed models, which is why I lean towards open models – even at the Mythos level – being preferable here if the capabilities exist in the first…
> it's not reasoning efficient enough so it doesn't show on UK AISIS's eval which is limited to 100M TOTAL tokens lmao bongs are all about efficiency… maybe too much in this case
very interesting convo between a current and ex researcher at anthropic also TIL kimi K3 might be mythos tier in terms of cyber. it's not reasoning efficient enough so it doesn't show on UK AISIS's eval which is limited to 100M TOTAL tokens, which is kinda crazy imo, seems very…