Two Bugs Took Over OpenAI Employee ChatGPT and Codex Accounts

@S1r1u5_’s team chained a libheif overflow and an SSO flaw into account takeover in under 72 hours.

On July 25 the Hacktron team, led by @rootxharsh with @S1r1u5_ and @iamnoooob, took over ChatGPT and Codex accounts belonging to OpenAI employees. The chain started with a HEIC upload to community.openai.com: a known libheif heap overflow, fixed upstream but never marked security-relevant, still present in Discourse, gave them remote code execution on the forum.

The second bug was an OpenAI SSO flaw that turned the forum foothold into employee ChatGPT and Codex accounts, some connected to Outlook, Slack, GitHub, and Google Drive. To prove impact without reading sensitive code, they had Codex open a harmless pull request in OpenAI’s internal monorepo. OpenAI fixed the SSO issue roughly 14 hours after the report and paid a $6,500 bounty.

@S1r1u5_ credits AI agents with a meaningful share of the exploit work: Opus 4.8 found the libheif vulnerability and built a partial exploit, and Opus 5 adapted it to Discourse hours after launch. The same team’s wider disclosure, HEIF Heist, puts the library beneath Slack, Meta, GitHub Enterprise, Rails, and Next.js, with only one organization noticing mid-exploitation. “AI is reducing the amount of scarce expertise needed to develop exploits,” he writes. Work that once took months now takes days.

From posts by @S1r1u5_, @rootxharsh

Agents & Coding

A $50 API Account, Blender, and a 3D Printer

@jmacftw prints physical parts he never modeled himself, driven by DeepSeek 4.1 Flash through DSH.

@jmacftw’s recipe: a sub-$400 printer, $50 in a DeepSeek API account, DSH (DeepSeek Harness) running 4.1 Flash, and free Blender wired to an MCP server so the model drives it. He prints fit-check sections of fiddly parts before committing, and sends photos back when something fails. A rack on his desk “didn’t exist until a couple days ago,” and he never touched Blender himself.

From @jmacftw

Models & Hardware

Everything That Runs on One DGX Spark Right Now

@yume_arasaki catalogs a 180B MoE, a 300B-class vision model, video, voice, and music on a single box.

@yume_arasaki reports Qwen3.8-Flash-Next at 48.7 tok/s decode on one Spark, a figure he reproduced at 48.6, with 116 tok/s at 4 concurrent streams. He credits recipes that move the model’s 51B lookup table onto the SSD. Cruz packed DeepSeek-V4-Flash-Vision into about 95 GiB at 19.7 tok/s, and GLM-5.3-Flash into about 91 GiB at 24.6, both 300B-class MoEs on one box. MiniMax H3 turns out 5 seconds of 480p video with audio in about 68 seconds using NVIDIA’s Sol engine.

From @yume_arasaki

Same Gear Prompt, Two Flash Models, Two Very Different Runs

@WescheNex1q’s gear test: DeepSeek V4.1 Flash runs faster; GLM 5.3 Flash verifies.

@WescheNex1q asked each model for three meshing gears, 12, 20, and 32 teeth, over 240 frames, one shot with thinking on. DeepSeek V4.1 Flash TP4 finished in 13.5 minutes at 57 tok/s with perfect kinematics, but its self-check tested the gears un-rotated, read a 240/240 failure, and shipped anyway. GLM-5.3 Flash TP2 took 64 minutes at 24 tok/s, verified backlash at 0.004 across 960 phases, and finished materials.

From @WescheNex1q

GLM 5.3 Flash Is the Pick for a 2x Spark Setup

@landontgreen calls it the best model to run on two Sparks, as fast as Claude or ChatGPT.

@landontgreen reports GLM 5.3 Flash is such a good experience that he calls it the best model to run on a 2x DGX Spark setup, feeling just as fast, and at times faster, than his Claude or ChatGPT use. In the replies, @OneCloudEmoji uses it for decompiling work but keeps ds4 vision for speed on everything else.

From @landontgreen

Sources

Thursday, September 17, 2026 Back to the Archive Saturday, September 19, 2026