Calif CEO Thai Duong argues AI currently helps attackers more than defenders, since attackers are always early adopters of new technology.
Calif's team used OpenAI's Pass the Planet program to find PHP bugs similar to one Stefan Esser disclosed 22 years ago, after the PHP Foundation declined to fix a bug it had found two years earlier.
Thai Duong says AI companies sell tokens largely used for coding, so an agent good at coding is basically a hacker and simple fixes will not work.
AI Turns Months of Hacking Into Days, and Attackers Benefit
AI has turned finding and exploiting software flaws from slow specialist work into something one person with a model can do in days. Calif's hackers say this currently favors attackers, because they adopt new tools first and defenders lag. The risk cannot be trained out of models, since any agent good at coding is good at hacking, and long-trusted software like PHP and Chrome is proving full of exploitable bugs.
Takeaways
AI compresses exploit work from months to days. Bruce Dang says work that took many people three months now takes one person a few days. Calif built a full iPhone exploit chain in days with one engineer, and specialists spent about $5 on an open source model to reverse engineer printer firmware.
Mature software hides far more bugs than assumed. Thai Duong says bug counts came in about 10 times higher than expected, with roughly a thousand vulnerabilities found in Chrome. Microsoft's monthly updates have climbed from about 20 to about 1,400 as AI speeds identification.
Patch-to-exploit window has shrunk to hours. Publishing a patch now exposes the bug before users can upgrade, a category Thai calls minus-one-day vulnerabilities. He is sure tens of people watch PHP today, ready to exploit any newly disclosed flaw.
The guests split on who AI favors. Bruce thinks defenders can use AI to build systems that are hard to exploit and eventually counterbalance attackers. Thai says attackers are always early adopters and AI currently helps them more, though he believes good guys outnumber bad guys.
A capable coding agent is inherently a hacker. Thai rejects Jensen Huang's view that models can simply be fixed so they do not hack. AI companies sell tokens largely used for coding, and any agent good at coding cannot avoid being good at hacking.
Detection alone fails once intruders move fast. At two bank clients, an intruder who entered on Monday was everywhere by Tuesday. OpenAI's responders saw about a hundred hackers inside its network at once after a Hugging Face incident, so Calif no longer treats detection as enough.
Formal verification cannot cover today's codebases. Verification works for small systems such as a verified operating system core, but applications on top still have issues. Browsers run to billions of lines, and Thai says software changes too easily to verify fully without AI help.
How Calif cracked a 22-year-old PHP bug pattern
Calif joined OpenAI's Pass the Planet program, which funds firms to help open source maintainers, and Stefan Esser introduced the team to the PHP Foundation.
Tap a name to go there · tap a line to see why they are linked
On this episode 4
Show 1
Ryan Osteen, Calif's PHP historian, searched for bugs resembling one Esser disclosed 22 years ago and found many.
The PHP Foundation had found the same bug two years earlier but left it unfixed because the fix was expensive, risked new bugs, and nobody had shown a working exploit.
Calif delivered a working exploit the next day, with AI turning exploitation into an engineering problem.
How long humans stay in the loop
Calif believes hacking has entered a centaur era where the best hacker combines AI with a human expert.
In chess, Deep Blue beat Garry Kasparov in 1997 and centaur play lasted about ten years.
In Go, AlphaGo beat Lee Sedol about 20 years after Deep Blue, and a program with no human input arrived about three years later.
Thai says the human role may last three months or three years. Joe presumes it will not last long.