OpenAI published a blog post about an “autonomous vulnerability detection and remediation pipeline” that spawns countless AI agents that not only detect a huge number of software vulnerabilities, but also create code to fix them.
OpenAI published a blog post about an “autonomous vulnerability detection and remediation pipeline” that spawns countless AI agents that not only detect a huge number of software vulnerabilities, but also create code to fix them.
At the same time, according to OpenAI, the cost of the API, which concerns the AI tokens used in all the “thinking” processes within a single launch of this giant workflow, will be over $1.2 million.
“It’s a bit like a giant, multi-armed, automated colonoscopy for a very complex code base that doesn’t just look for dangerous polyps, but essentially removes them,” Gizmodo journalists commented .
Meanwhile, OpenAI claims that this particular system, designed to scan the Linux kernel, found 2,308 potential vulnerabilities. After further analysis, the system fixed the most critical of these issues by running those processes in a virtual machine. In 948 cases, OpenAI says, the vulnerabilities turned out to be truly dangerous.
“The agents then wrote the 948 necessary patches, but did not actually edit the code base. The patches were first sent to people for approval,” the publication explains.
So, if you round this up, the total of over $1.2 million in tokens is approximately $1,200 for each verified exploit.
Since such work is usually done by paying "white hat" hackers a bounty for finding bugs in the amount of $20,000 to $70,000, OpenAI is confident that spending $1.2 million in tokens is actually a very profitable investment.
“Applying these reward amounts to 948 exploits yields a hypothetical estimate of $18.9–$66.3 million, which is approximately 16–55 times the cost of our pipeline,” they say.
Previously, OpenAI began limiting the number of paid ChatGPT accounts on one device and subscriptions paid for with one card.
Additionally, OpenAI will merge the ChatGPT and Codex limits, so developers will have to share tokens between the two tools.
