I Found a $90k Remote Code Execution Vulnerability in Google

Original Article ↗ Hacker News Discussion ↗

A $90k remote code execution bounty awarded by Google has sparked debate over how much credit should go to the human researcher versus the AI assistant reportedly used to help find and describe the vulnerability. Commenters argue over whether large language models are “just tools” like nmap or Photoshop, or qualitatively different systems that warrant explicit attribution when they shape both the technical work and the write-up. The thread also explores the future of bug bounties in an AI-assisted world, including concerns about shrinking payouts, legal versus black-market incentives, and how AI may widen the gap between those with and without access to powerful models.

Use of LLMs in vulnerability research

  • Many comments assume an LLM was heavily used to find or explain the bug.
  • Several argue LLMs are “just tools” like scanners or debuggers; the human still needs domain knowledge, strategy, and effort.
  • Others contend that if the model did the key reasoning or discovery, it deserves explicit credit; otherwise the human is overstating their role.
  • Some note that getting valuable results from current models still requires substantial skill and guidance.

Authorship, attribution, and disclosure

  • Multiple commenters find the write-up strongly “LLM-flavored” and dislike first‑person narration if text was largely machine‑generated.
  • A recurring view: using LLMs is fine, but not disclosing heavy use—especially for a first‑person story—is deceptive.
  • Counterpoint: we don’t attribute every other tool (e.g., nmap), and insisting on AI disclosure for everything is inconsistent.
  • One nuanced position: disclosure is less important than clearly demonstrating genuine human understanding in the write-up.

Bug bounty economics and ethics

  • Debate over why bounties exist:
    • One side: to make reporting more attractive than selling exploits or abusing them.
    • Another: for prioritizing engineering work, since users rarely report security flaws.
  • Disagreement over black‑market value of this RCE; some say sub‑$1,000, others think it would be worth far more to attackers.
  • Several expect AI to push bounty prices down by making vulnerability discovery cheaper, especially for web vulns.

Access, inequality, and AI “tool” status

  • Concern that access to powerful models sometimes requires applications and ID verification, potentially creating a privileged “club.”
  • Some foresee a future divide: people with access to advanced models gaining disproportionate economic advantage.
  • There’s a side debate over whether AI is “just a tool” or fundamentally different from traditional tools, given how much “heavy lifting” it can do.

Technical and process commentary

  • Some are surprised such a basic‑seeming issue (pickle, key in cookie) survived reviews; others emphasize that all large systems have bugs.
  • A few note orgs that ban untrusted pickle entirely, likening it to eval.
  • Mixed reactions to the write-up quality: impressive find, but many find the LLM-style prose off‑putting and say they now skim or skip such articles.