"This is the essence of the problem. Attack and defense are asymmetric; the defender needs to defend across their entire attack surface, while the attacker needs to find just one flaw. However, the number of flaws is limited; once you've found most of them, finding more is hard. If you've found..."
...all of them, it doesn't matter how smart the attacker is, they will not be able to invent more from thin air. To defend effectively against attacks, people writing software need to have access to good models, without restrictions, to check over their work and make sure that it does not have bugs in it. Delaying or impeding their access just gives an attacker, who probably has no impediments to their own access, the ability to find flaws that the defender doesn't have the capacity to find first. — Perry E. Metzger
Source: https://x.com/perrymetzger/status/2079260327972065582
The cybersecurity debate on open-source AI is backwards. Open models aren't the risk, they're the defense! Attackers can already jailbreak any API or guardrails. Defenders can't secure systems with black boxes they can't control, inspect, test, or run locally. — clem
Source: https://x.com/ClementDelangue/status/2079253659108409587