Hacktron researchers chained a libheif heap-buffer-overflow in Discourse’s image pipeline with an OpenAI SSO misconfiguration to take over multiple OpenAI employee ChatGPT/Codex accounts and reach internal repositories. Image uploads bypassed FastImage and invoked ImageMagick, exposing vulnerable libheif versions (Debian 12’s 1.19.7 and Debian 13’s 1.19.8) that lacked a backported security fix. Using an exploit developed against Discourse’s Docker image, the team gained RCE, then leveraged the SSO flaw to access an employee Codex and produce a harmless pull request in OpenAI’s internal monorepo as proof. The chain was discovered and escalated in under 72 hours; Discourse deployed fixes and sandboxing, Debian published an update on Aug 8, 2026, and OpenAI awarded a $6,500 bounty.
The write-up details how modern LLMs accelerated exploit development: Claude Opus 4.8 began the work but Opus 5 and later GPT-5 Sol rapidly produced reliable ARM64 and x86-64 exploits (including adjustments for jemalloc and ASLR). The broader “HEIF Heist” effort found libheif widely embedded across major platforms, meaning many image-processing pipelines are at risk. The campaign cost under $3,000 in model tokens and only days of human time per target. Recommendations include rebuilding Discourse Docker images, sandboxing ImageMagick, and auditing services that accept .heic/.heif/.avif uploads.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
ZCode, an AI coding app from Z.ai, silently uploads users' entire Git histories to Alibaba Cloud, encrypting the data with a key only Z.ai holds. The app's closed-source nature and data handling practices raise privacy concerns among developers and security experts. (tokenstead.ai)
Google is facing a legal challenge over allegations of illegally retaining customer data. The case raises issues about data privacy and Google's data management practices. (medium.com)
A Microsoft executive described AI scraping as 'the largest theft of labor in human history,' according to unredacted filings in a lawsuit. The filings reveal that companies like OpenAI bypassed paywalls and stripped copyright notices to build AI training datasets, raising legal and ethical concerns. (techcrunch.com)
Researchers have developed a method to eavesdrop on audio played through headphones from a distance. This vulnerability could allow attackers to listen in on private conversations without direct access to the device. (theregister.com)
ZCode, an AI coding desktop app, silently encrypts and uploads users' entire workspace, including Git history, to the cloud without clear user consent. The encrypted data is stored on Aliyun OSS, with the private key kept in the cloud, preventing decryption locally. (blog.ferstar.org)
Emulating x86 memory behavior on ARM hardware faces significant challenges due to differences between the strict x86-TSO model and ARM's relaxed memory model. These differences impact how memory operations are ordered and visible across processors, complicating software emulation efforts. (fex-emu.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.