The Nightly Build.
The Wire

GPT-5.6 Codex deleted users' home directories, and OpenAI's model card had flagged the risk

OpenAI's Codex lead confirmed GPT-5.6 can erase a user's home directory in full-access mode, a behavior its own model card had already rated a severity-3 misalignment risk.

4 min read2026-07-17
engineering

GPT-5.6 Codex deleted users' home directories, and OpenAI's model card had flagged the risk1

Never hand a coding agent full-access mode without a sandbox: the OpenAI Codex lead has now confirmed GPT-5.6 can erase a user's home directory when it runs that way. The mechanism is a path-construction slip, not sabotage. Tibo Sottiaux described the model as attempting to "override the $HOME environment variable to designate a temporary directory," then accidentally deleting "$HOME itself," and said it happens "specifically when running in full-access mode without sandboxing or auto-review protections."2 The reports name real losses: investor Matt Shumer said the model "just accidentally deleted almost ALL of my Mac's files," and engineer Bruno Lemos lost an entire production database and said it had "never happened to me before, with any other model, ever."3 The disclosure that reframes the incident is OpenAI's own: its GPT-5.6 model card had already documented the behavior as a "severity level 3" misaligned action, appearing "a bit more often in misalignment simulations" than GPT-5.5.3 The deletions were a known, pre-flagged risk rather than a surprise. OpenAI's guardrail in the changelog is a warning, not a wall: it "Added clearer Full access warnings and dialog when combining Full access with Ultra."1 Techzine reads file deletion as "a well-known phenomenon" for LLMs and prescribes the mitigation the vendor note underplays, "robust file backups when granting generous AI permissions."4 The sandbox and auto-review are the guardrail; the dialog box is not.

industry

Kimi K3 claims the first open 3-trillion model, but the weights stay locked until July 275

The open-weights reflex, grab it and self-host cheap, does not apply to Kimi K3 yet, and the price no longer makes it a bargain. Moonshot announced K3 as the "world's first open 3T-class model," a 2.8-trillion-parameter model built on its "Kimi Delta Attention and Attention Residuals" with a Stable LatentMoE "effectively activating 16 out of 896 experts" and a 1M-token context.5 A mixture-of-experts model routes each token through a few of many expert subnetworks, so only a slice of those 2.8 trillion parameters runs per token. Two caveats the launch tweet dropped are the builder story. First, you cannot self-host it: Moonshot says "the full model weights will be released by July 27, 2026," and until then K3 is API-only.5 Second, the price climbed a tier. K3 lists at "$3.00/MTok for cache-miss input" and "$15.00/MTok for output" per million tokens, with $0.30 cache-hit input.5 Simon Willison puts the prior Kimi K2.6 at roughly $0.95 in and $4 out, so cache-miss input has about tripled and output climbed roughly threefold to fourfold.6 K3 now sits at Anthropic's Sonnet tier, so the open-weights bargain is gone. Moonshot's self-reported wins run against a mid tier, and it concedes K3 still "trails Claude Fable 5 and GPT 5.6 Sol."5 Willison adds third-party figures the lab omits: Artificial Analysis scored K3 at an "overall Elo of 1547," and it leads Arena.ai's Frontend Code arena.6

research

Thinking Machines' Inkling is a 975B open model its own lab says will not win7

Inkling is built to be fine-tuned into your task, not to top a leaderboard, and the lab says so in the release. Thinking Machines Lab introduced it as an Apache-2.0 model that "reasons natively over text, images, and audio," with a context window "of up to 1M tokens" and pretraining on "45 trillion tokens."7 The number that governs your bill is the split: it is a sparse mixture-of-experts model with 975B total parameters but only 41B active, so serving cost tracks the 41B, closer to a 41B dense model than to the 975B headline.7 The lab disowns the leaderboard outright, stating "Inkling is not the strongest overall model available today, open or closed," and positioning it as "a good open-weights base for customization" with fine-tuning on its Tinker service.7 A smaller sibling, Inkling-Small, is 276B total and 12B active. On Tinker the model is live "with context length options of 64K and 256K tokens" at "a 50% discount for a limited time."7 Willison places it in the U.S. open-weights field as "comparable to NVIDIA Nemotron and Gemma 4," a peer-set the lab's own post leaves out.8

compute

LongStraw runs million-token RL post-training on a fixed GPU budget, no new hardware9

LongStraw lets RL post-training run at the million-token context a deployed agent actually sees, on a fixed GPU budget, instead of training short and trusting length generalization to hold at deployment. Zhou et al. report that inference is "approaching million-token contexts, while post-training workloads often remain at 256K tokens or below" — the gap that bites agents hardest, whose observations, tool outputs, and prior decisions pile up over long trajectories. It is not a new attention kernel but a GRPO-based execution stack that "evaluates the shared prompt without autograd, retains only model-specific state needed by later tokens, and replays short response branches one at a time," shrinking the live training graph "at the cost of additional replay time." That is a time-for-memory trade you opt into, not a free one. In the authors' self-reported measurements, 8× H20 GPUs reach 2.1M positions for groups of 2 and 8, each larger group adding only 0.21 GB of peak allocated memory, with a stress test at 4.46M; 32× H20 validates a full 2.1M-token prompt across all 78 layers of GLM-5.2. H20 is a deliberately memory- and bandwidth-limited part, so the roughly eightfold move from 256K to 2.1M comes with no new hardware.9

Sources

  1. OpenAI · Codex changelog
  2. Tibo Sottiaux (OpenAI Codex lead) · statement on Codex file deletions
  3. The Register · OpenAI admits GPT-5.6 occasionally deletes files, but it's an "honest mistake"
  4. Techzine · OpenAI explains why GPT-5.6 Sol deletes files
  5. Moonshot · Kimi K3
  6. Simon Willison · Kimi K3, and what we can still learn from the pelican benchmark
  7. Thinking Machines Lab · Introducing Inkling
  8. Simon Willison · Inkling: Open-weights model from Thinking Machines Lab
  9. arXiv · Zhou et al., "LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget"