Three researchers used Claude to reach OpenAI's internal GitHub in 72 hours for a $6,500 bounty

verifiedDeveloperLegal

Friday, September 18, 2026

Confidence

High · — WSJ primary + Forbes, Invezz, Tech Portal corroborating; downstream impact of source-code access unknown

Evidence

bug-bounty disclosure + WSJ reporting + Hacktron writeup

Hacktron AI ran Claude in an autonomous loop against an OpenAI-side Discourse flaw, chained SSO tokens into an employee's Codex, and opened a pull request inside the internal openai/openai monorepo.

The details

Under $3,000 in tokens, $6,500 bounty; WSJ broke it, OpenAI patched in ~14 hours.
The team used a Claude version restricted to qualified cybersecurity practitioners.

What it means

A cybersecurity-tier Claude built for defenders was the load-bearing tool in an offensive chain against a rival lab — AI-vendor liability is no longer theoretical. Every red-team budget just re-baselined against three people and API credits.

Hacktron AI ran Claude in an autonomous loop against an OpenAI-side Discourse flaw, chained SSO tokens into an employee's Codex, and opened a pull request inside the internal openai/openai monorepo.

Under $3,000 in tokens, $6,500 bounty; WSJ broke it, OpenAI patched in ~14 hours.
The team used a Claude version restricted to qualified cybersecurity practitioners.

Sources

Apply this today

The hands-on layer the brief points to: a workflow you can run, a tool to test, and today’s 60-second video.

Workflow of the day

Draft an AI Vendor Risk Brief from a CEO's Public Statements

Legal · tested · saves 60-90 minutes per vendor review

Full walkthrough →

Tool watch

Comp AI

Comp AI's continuous compliance play: does it change how you scope a SOC 2 engagement?

Read the breakdown →

▶ Watch today’s 60-second brief →

In this issue

← Back to brief

Get the next one in your inbox

One daily brief. Every story gets a hype verdict.

No spam. Unsubscribe anytime.

On this page

The details
What it means

In this issue

← Back to brief

Stat of the Day

–93%▼

agentic-payments rail volume, 2026

American Banker reports a specialized agentic-payments rail has lost 93% of its volume this year even as startups pile into the category — the rare demand-side number in an infrastructure-heavy narrative. Source.

Today’s Take

Today's slate reads as one thesis in four beats: labs are vertically integrating (Astra, Anthropic biology), the tools are already offensive against each other (Claude on OpenAI), the deployment claims are outrunning the disclosures (Ant International), and the buyers keep saying the agents are not ready to run alone (WSJ, AT&T, Crocs). The pattern is a supply side sprinting past a trust ceiling — and a rare demand number (–93%) that suggests the ceiling is real, not rhetorical. The near-term bet that works is boring: narrow, audited, human-supervised agents inside regulated workflows. Considered and passed: SoftBank's $25B Arm margin loan (Bloomberg primary but a leverage-mechanics story more than an agent story) and the BoE governor's financial-crime AI remarks (trade-press intermediary, no BoE primary transcript).

— Agentic desk

Never miss a signal

Join professionals getting the daily AI brief every weekday morning.

No spam · Unsubscribe anytime

No sponsored verdicts
Sources linked
Corrections public