OpenAI quadruples bug bounties while open source fixers run short

OpenAI said on 3 November that it has quadrupled bug bounty payouts, with critical flaws in ChatGPT or Codex now worth up to $30,000 and exceptional finds reaching about $172,000 once bonuses are counted. The open source projects underneath it cannot match that pace.

For anyone running a website or a dependency tree, the bottleneck has moved. Finding bugs is getting cheaper and faster. Validating them, writing patches and shipping releases is not, and nobody has published a plan to fund that half of the job.

The OpenAI programme, reported by The Decoder on 3 November, pays up to $100,000 for exceptional findings, with quality, loyalty and campaign bonuses on top. A five-tier loyalty scheme and time-limited campaigns tied to product launches are new, and researchers who reported a flaw in the previous 30 days get the difference backdated. OpenAI is also preparing a security sprint with the firm Cathedral. During the last one, 25% of OpenAI's product engineers worked on finding vulnerabilities.

Paying outsiders is one lever. Staffing is another, and there the numbers run the other way.

Anthropic's free OSS Scanner, which opened enrollment on 8 October, gives eligible open source projects recurring code reviews from its frontier models, including Claude Mythos. Each report carries a reproducer, an explanation and, where available, a candidate patch. The output is entirely model-generated, with no human triage. Anthropic says reports may be wrong. TechTarget reported on 30 October that the models had turned up 29,432 candidate vulnerabilities. Anthropic had manually reviewed 6,123 of them, and 513 findings were patched upstream. When pen testers checked 97 critical and high severity findings across 48 projects, 88% met the bar for disclosure; 11 were duplicates and one was a false positive.

"The scan is free, but the time needed to deal with its findings isn't," Krishna Shakula, an AI engineer at the law firm Robertson, Anschutz, Schneid, Crane & Partners, told TechTarget.

That gap shows up in the work itself. A model can flag a suspicious pattern in seconds. A maintainer still has to reproduce it, decide whether it is exploitable in that project, judge urgency and check the proposed fix. Even a genuine vulnerability can take hours to resolve safely.

The arithmetic is starkest in PHP. The scripting language sits behind roughly 69% of sites whose server-side language is known, according to W3Techs estimates cited by BigGo Finance. The PHP Foundation, which coordinates security responses for that code base, has one full-time employee. Executive director Elizabeth Barron told BigGo Finance she works with about 14 part-time contractors and unpaid board members. Demand for fixes has doubled since AI made flaw hunting easier, she said, but the number of collaborators has not. Two critical PHP flaws found by a team using OpenAI models forced emergency patches, and Cloudflare said the impact reached servers running PHP 5.4 and later.

Cloudflare chief executive Matthew Prince said in a post on X that millions of attack attempts were observed within 24 hours of the disclosure. The FBI's cyber division warned the flaws could allow remote code execution in the PHP core.

Cloudflare added firewall rules to block the attacks, and GoDaddy patched its managed hosting and told customers they need not act. Cloudflare also opened a capture-the-flag contest on HackerOne, with prizes from $100 to $10,000, built around a deliberately vulnerable PHP environment. Awareness is thinner at the small end. In a Network Solutions survey of 121 US small business owners with their own sites, 45% had never heard of PHP.

Free defensive tooling keeps arriving. The PrivacySafe Foundation published 57 open source scripts covering 139 flaws across 67 projects and 10 programming languages, Security Today reported on 29 October. The list includes PHP2Shell, and most scripts audit configurations without changing system code.

On 3 November a free, open source AI red team course went up with more than 60 hours of material and Docker labs built on garak, PyRIT and promptfoo. It teaches people to break chatbots and backdoor models. The skills are the same ones maintainers need to check whether a machine-generated finding is real.

Money is not flowing to that side of the ledger. OpenAI's bounty pays for findings about OpenAI's own products, and Anthropic's scanner pays nothing for the patches it suggests. Barron says funding to support PHP is severely lacking.

Neither OpenAI nor Anthropic has published how many projects its programme covers. The size of the gap is still guesswork.