July 15, 2026 (Wed)

[01]

Top Stories

1. Grok Build CLI secretly uploaded entire Git repos — 27,800× excess data, .env secrets included

Security researcher cereblab published wire-level analysis showing Grok Build CLI v0.2.93 transmitted 5.1 GiB from an 11.2 GiB repo when only 192 KiB was needed
Full Git history, untracked files, and unredacted .env secrets (API keys, DB passwords, SSH keys) sent to xAI's Google Cloud Storage bucket grok-code-session-traces
xAI had marketed the tool as: "nothing from your codebase transmitted to xAI servers during a session" — wire data directly contradicts this
The /privacy toggle was ineffective: it only controls post-upload retention, not whether upload happens
Story hit Hacker News front page July 14; xAI deployed a server-side fix (disable_codebase_upload: true) ~24h after disclosure
Elon Musk pledged global deletion of all previously uploaded user data (July 13 on X) — not independently verified
Why it matters: the highest-signal trust failure in the agentic coding CLI space since the category emerged. Every developer who ran Grok Build before July 13 on a repo containing credentials should rotate them now. Sets back Grok Build's enterprise adoption at the moment xAI is competing hardest with Claude Code.
→ xAI · Grok Build (source) (The Hacker News)

2. Gemini 3.5 Pro July 17 GA confirmed — full architectural rebuild complete, 2M context

Multiple sources as of July 14 confirm tomorrow (July 17) as the firm GA date for Gemini 3.5 Pro — superseding the earlier July 22-28 slip estimate
The full architectural rebuild of the 2.5 Pro base is complete; the rewrite targeted math reasoning, SVG/image quality, and long-horizon coding
Leaked specs (not yet official): 2M token context window, Deep Think Reasoning Layer (gated), autonomous workflow mode
No official benchmark sheet released yet; full data expected at GA announcement
Simultaneously: Google has capped Meta's Gemini compute access when capacity tightened — Meta's AI projects delayed; seen as Google's vertical integration advantage (owns models + TPUs + cloud)
Why it matters: the architectural rebuild is the first time a major frontier lab has scrapped and restarted a pre-training cycle mid-release cycle. If the rebuild delivers, Gemini 3.5 Pro could shift the competitive balance at the precise moment Fable 5 is credits-only and Grok Build is reeling from the privacy breach.
→ Gemini 3.5 Pro (source)

3. Anthropic Claude for Teachers — free premium access for US K-12 educators

Anthropic launched Claude for Teachers on July 14: verified US K-12 teachers get one free year of premium Claude (sign up by June 30, 2027), including Claude Code, Claude Cowork, and a Learning Commons connector with standards for all 50 states
Teaching skills library co-developed with learning scientists; FERPA-compliant K-12 data agreement; training data excluded
Ecosystem: ASSISTments, Brisk Teaching, Canva Education, Coteach, Diffit, Eedi, MagicSchool, Snorkl
Why it matters: US public school systems require FERPA compliance and training-data exclusion before granting IT approval — Anthropic has cleared both bars, positioning Claude ahead of Microsoft Copilot for Education and Google Workspace for Education. Reach: ~3.7M US K-12 teachers. The education layer could feed enterprise familiarity with Claude at scale.
→ Anthropic (source) (Anthropic)

[02]

Long-Horizon-Terminal-Bench (LHTB) — arXiv:2607.08964

HF Daily #1 paper July 14 (25 stars)
46-task benchmark measuring LLM agent performance in a stateful containerized terminal over hundreds of steps, graded with dense reward (rebuild-from-artifact verifiers — self-reported progress doesn't count)
SOTA result: best model achieves only 15.2% pass@1 at 0.95 partial-reward threshold; perfect completion (1.0) only 10.9%
Average token consumption: 9.9 million per task; average execution: 85.3 minutes per run
Why read it: the evaluation methodology (dense reward + containerized stateful env) is a step-change from binary pass/fail short-task benchmarks — and the 15% ceiling tells you exactly how far current "agentic" systems are from reliable long-horizon autonomy
→ Long-Horizon-Terminal-Bench (LHTB)

[03]

Gemini 3.5 Pro GA: tomorrow (July 17) — official benchmark sheet and pricing expected; may finally shift the leaderboard from the Fable 5-suspended, GPT-5.6-dominant landscape. Watch for Deep Think benchmark vs. Fable 5 SWE-bench Pro 80.3%.
Grok Build credential rotation: if you or your team used Grok Build CLI before July 13, 2026 — rotate API keys, DB passwords, SSH keys, and any secrets that were in tracked files or Git history. xAI's deletion pledge is not independently verified.

[04]

Long-Horizon-Terminal-Bench (LHTB) (new — long-horizon terminal agent benchmark; HF Daily #1 July 14; SOTA 15.2%)

[05]

xAI: Grok Build full-repo privacy breach (July 12-14); tags updated
Grok Build: Privacy Breach section added with full technical detail and security advisory
Anthropic: Claude for Teachers (July 14) added to Recent Activity
Gemini 3.5 Pro: Status updated — July 17 GA re-confirmed (supersedes July 22-28 slip); timeline entry added; source updated