[ad_1]
OpenAI has launched a bug bounty, encouraging members of the general public to seek out and disclose vulnerabilities in its AI companies together with ChatGPT. Rewards vary from $200 for “low-severity findings” to $20,000 for “distinctive discoveries,” and stories are submittable through crowdsourcing cybersecurity platform Bugcrowd.
Notably, the bounty excludes rewards for jailbreaking ChatGPT or inflicting it to generate malicious code or textual content. “Points associated to the content material of mannequin prompts and responses are strictly out of scope, and won’t be rewarded,” says OpenAI’s Bugcrowd page.
Jailbreaking ChatGPT often entails inputting elaborate eventualities within the system that permit it to bypass its personal security filters. These may embody encouraging the chatbot to roleplay as its “evil twin,” letting the person elicit in any other case banned responses, like hate speech or directions for making weapons.
OpenAI says that such “mannequin issues of safety don’t match effectively inside a bug bounty program, as they aren’t particular person, discrete bugs that may be instantly fastened.” The corporate notes that “addressing these points usually entails substantial analysis and a broader strategy” and stories for such issues needs to be submitted through the corporate’s model feedback page.
Though such jailbreaks reveal the broader vulnerabilities of AI techniques, they’re possible much less of an issue instantly for OpenAI in comparison with conventional safety failures. For instance, final month, a hacker referred to as rez0 was capable of reveal 80 “secret plugins” for the ChatGPT API — as-yet-unreleased or experimental add-ons for the corporate’s chatbot. (Rez0 famous that the vulnerability was patched inside a day after they disclosed it on Twitter.)
As one person replied to the tweet thread: “In the event that they solely had a paid #BugBounty program – I’m sure the group might assist them catch these edge-cases sooner or later : )”
[ad_2]
Source link