Digest · 10 stories · 22 sources

2026-07-13

Updated

Suggest a story

Top stories

  1. Security developing xAI Grok Build CLI uploads the entire repository and unredacted .env secrets by default A researcher published a mitmproxy wire capture of xAI's Grok Build CLI (grok 0.2.93) reporting that it uploads the full working repository, every tracked file plus complete git history, to a Google Cloud Storage bucket named grok-code-session-traces through a POST /v1/storage channel, independent of what the agent reads. On a 12 GB test repository the storage channel moved 5.10 GiB across about 73 chunks while the model-turn channel carried 192 KB, and planted files the agent never opened were recovered verbatim from the uploaded git bundle. Contents of a .env file, including canary APIKEY and DBPASSWORD values, appeared unredacted in both the POST /v1/responses bodies and a session-state archive. Disabling the "Improve the model" toggle did not stop the upload, and /v1/settings still returned traceuploadenabled: true. The author states the capture does not prove xAI trains on the data. A separate user account raised on Hacker News later on 2026-07-13 claims the CLI uploaded the entire home directory, not only the working repository, which would widen the reported scope. That claim comes from a single social-media account and is not independently verified.
  2. Agentic coding discussion Claude Code sends about 33k tokens of overhead before the prompt, OpenCode about 7k A logging-proxy measurement published 2026-07-12 reports that Claude Code 2.1.207 sends roughly 33,000 tokens of system prompt, tool schemas, and injected reminder blocks before the user prompt on claude-sonnet-4-5, against about 7,000 for OpenCode 1.17.18, a 4.7x floor difference. The gap is mostly tool schemas (27 tools, about 24,000 tokens, versus 10 tools, about 4,800 tokens) plus a larger system prompt (about 6,500 versus about 2,000 tokens). The author reports the ratio narrows to about 3.3x on Claude Fable 5 and that Claude Code rewrites its cached prefix mid-session, writing up to 54x more cache tokens than OpenCode's byte-identical prefix.
  3. Security confirmed Chromium 148 makes Math.tanh a browser fingerprint of the host OS Since Chromium 148 (V8 14.8.57, commit c1486295ae5), Math.tanh calls the host operating system math library (std::tanh) instead of the bundled fdlibm implementation, so the last-bit result now varies by OS. Math.tanh(0.8) returns distinct values on Linux (glibc libm), macOS (libsystemm), and Windows (UCRT ucrtbase.dll), giving JavaScript an OS signal independent of the user-agent string. IEEE 754 does not require correctly rounded transcendental functions, so each library uses different polynomial coefficients and produces results differing by about one unit in the last place.
  4. Agentic coding discussion Ploy reports a production agent moved to GPT-5.6 ran 2.2x faster at 27% lower cost Ploy reports migrating its production website-building agent from Claude Opus 4.8 to GPT-5.6 Sol and evaluating both against a fixture suite of hundreds of cases. Median wall-clock per build dropped from 8m00s to 3m42s (2.2x), cost from 3.06 to 2.22 USD (27%), and output tokens from 33.0K to 17.1K, with a small visual-quality gain. The switch required reworking the eval harness (tool-call budgets were Opus-specific), a "required but nullable" tool-schema transform because GPT-5.6 fills every parameter, a workspace-scoped prompt-cache rebuild reaching a 83.7% hit rate, and disabling server-side reasoning replay.

Conferences and events

  1. Event developing EuroPython 2026 is active in Kraków through 2026-07-19 EuroPython 2026 began 2026-07-13 at the ICE Kraków Congress Centre in Kraków, Poland, the 25th edition after three years in Prague. The main conference runs 2026-07-13 through 2026-07-17 with sprint days 2026-07-18 and 2026-07-19, covering CPython, the packaging and typing ecosystems, and scientific and web Python.

Security

  1. Security confirmed Unauthenticated RCE chain disclosed for the end-of-life Motorola MR2600 router A researcher published 2026-07-13 an unauthenticated remote code execution chain in the Motorola MR2600 router, whose last firmware (v1.0.22, mid-2024) is end-of-life. The chain combines improper SEAMA firmware-image validation in the upload endpoint, an authentication check that runs only after the malicious image is already written to /tmp/firmware.img, and inconsistent URI matching (substring allowlist but exact-match denylist) that bypasses auth through a crafted path such as a query-suffixed login URL. It is reachable from the LAN by default and remotely when remote management is enabled, with about 41 devices exposed at disclosure. Motorola Mobility and Motorola Solutions each disclaimed ownership and no fix was issued. No CVE is assigned.

Hacker News

  1. Pulse discussion Ask HN: add a flag for AI-generated articles The day's top thread (776 points) asks whether Hacker News should let users mark or flag AI-generated articles. Commenters question whether the existing voting and flagging already cover it, note that AI-detection tools are unreliable and any label would be contested, and link a parallel Lobsters discussion.
  2. Pulse discussion Zig creator's rebuttal reframes the Bun Rust rewrite as Anthropic marketing A 2026-07-12 opinion post by Ray Myers, the day's top thread at about 1,200 points, argues that Bun's Zig-to-Rust rewrite, done with Anthropic's Claude Code and cited by Anthropic as a Fable 5 showcase, was a marketing narrative more than a memory-safety necessity. Myers holds that Bun's memory bugs came from engineering practice rather than a Zig limitation, points to TigerBeetle as a reliable Zig project held together by a strict style, and reads the rewrite's own safety and readability choices as conceding that human judgment still matters. It backs Andrew Kelley's earlier technical rebuttal.
  3. Pulse discussion George Hotz argues the LLM technology is real and the hype is the problem An opinion post by George Hotz (2026-07-12) that reached the front page (369 points) separates the usefulness of large language models, which the author values highly as a coding and research tool, from the marketing framing around them, which the author criticizes as overstated. It is commentary with no primary release or benchmark and carries no verifiable engineering claim.
  4. Pulse discussion Show HN: Clawk runs coding agents in a disposable network-restricted Linux VM A Show HN project (122 points) gives a coding agent its own disposable Linux VM instead of the host machine. The user code is mounted in and the agent runs as root inside the guest with no permission prompts, while the host filesystem, keychain, and network stay out of reach behind a network allow-list that blocks outbound connections to unlisted servers. It is a single Go binary (Apache-2.0, macOS with experimental Linux support) that attaches Claude Code, Codex, or a shell, and it surfaced the same day as the Grok Build CLI exfiltration reports covered in Top stories.