AI #177 Part 1: Tip of the Iceberg
Zvi's AI Roundups · Zvi Mowshowitz · 2026-07-16
Zvi Mowshowitz's weekly AI roundup covers xAI's Grok secretly uploading private codebases, Apple suing OpenAI for systematic trade secret theft, the NYT accusing OpenAI of discovery fraud, and new model releases including GPT-5.6 Sol, Meta's Muse Spark 1.1, and Thinking Machines' Inkling.
Appears in
Extraction
Topics: ai-weekly-roundupxai-securityopenai-legalmodel-releasesai-safety-governance
Claims
- xAI's Grok Build binary was found to upload entire private Git repositories—including secrets preserved in git history—to xAI's cloud storage outside the tool-call permission system, even when tool calls were disabled.
- xAI silently rewrote its Frontier Artificial Intelligence Framework on June 30, 2026, removing both quantitative risk acceptance thresholds and whistleblower protections in favor of undefined qualitative language.
- Apple sued OpenAI alleging its hardware chief Tang Tan systematically directed former Apple employees to steal trade secrets during job interviews, with message transcripts cited as evidence and a broad preliminary injunction considered likely.
- The New York Times accused OpenAI of lying during copyright lawsuit discovery by concealing its ability to search training data and deleting output logs in violation of court preservation orders.
- Meta's Muse Spark 1.1 claims agentic and coding performance rivaling GPT-5.5 and Opus 4.8, but Zvi views the benchmark presentation skeptically given Meta's track record.
- OpenAI lost another safety head, with Johannes Heidecke departing and safety teams being reorganized under a combined research-and-safety VP.
Key quotes
How many super sus, shady and irresponsible things does xAI have to do, before we decide that we want nothing to do with their products even if they someday put out a good one?
This reads to me like OpenAI totally did what they are accused of doing. OpenAI doubtless thinks the whole lawsuit is stupid and the asks here are crazy and harmful, and I mostly think OpenAI is right about that, but if my read is right then this is not an acceptable way to handle such a situation.
The new FAIF removes the two quantitative risk acceptance criteria the framework contained... Both criteria are now gone. In their place, the new version adopts a qualitative 'systemic risk acceptance determination' borrowing the language of the EU General-Purpose AI Code of Practice.