Zynergy Wire
Sunday, 8 November 2026 · Private edition
粵語 0:00
Archive · Sunday, 8 November 2026 Back to Today All dates

Archive

Sunday, 8 November 2026

AI 6 items

The Hacker News (Nov 7, 2026) reports Apiiro research: when users pick “Yes, and don’t ask again,” Claude Code can save the exact approved shell command — including passwords and API keys — in .claude/settings.local.json. Soft title: AI Security Watch / coding-agent secret exposure with measured public GitHub impact — not a new CVE number and not a WorkOS/FMF re-card. Apiiro figures via THN: 104,185 public copies in 103,617 repos; 1,122 copies (~1.1%) hold ≥1 secret; 1,379 distinct secrets. Broader count incl. git history, shared .claude/settings.json, and agent transcripts: 2,539 secrets across 1,994 repos. More than a third of secrets open a database (379 database passwords); also 43 SSH logins, 42 GitHub/GitLab/npm tokens, 23 AWS keys; 63 organizations among owners. In 539 of 1,105 traced repos, a Claude co-author commit changed the file. One notified firm confirmed a live key; Apiiro did not test whether every secret still works. Global git-excludes don’t travel with the project / don’t undo commits; Claude Code does not add the file to project .gitignore. Of 16 coding agents compared, Claude Code is the only one confirmed to save the full approved command in a project file by default. Anthropic newsroom silent (no new Nov 7–8 article slug). Soft: Apiiro blog HTTP 500 this close — THN carries figures; live-exploit counts unknown.

Why it matters

Security Watch lead — measured public GitHub secret exposure from a default coding-agent “don’t ask again” settings file; rotate-first remediation stakes after weekend Sat WorkOS/FMF book.

The Hacker News (Apiiro research)
More ▾ Less ▴
Confirmed
What happened

The Hacker News on November 7, 2026 reports Apiiro research on Claude Code’s “Yes, and don’t ask again” path: the approved shell command — including passwords and API keys — can be saved verbatim into .claude/settings.local.json. Soft title: AI Security Watch / coding-agent secret exposure with measured public GitHub impact — not a new CVE number and not a WorkOS/FMF re-card. Stamp: THN Nov 7, 2026 (after Sat LIVE). Apiiro figures as carried by THN: 104,185 public copies across 103,617 repos; 1,122 copies (about 1.1%) hold at least one secret; 1,379 distinct secrets. A broader count that includes git history, shared .claude/settings.json, and agent transcripts finds 2,539 secrets across 1,994 repos. More than a third of secrets open a database (379 database passwords); also 43 SSH logins, 42 GitHub/GitLab/npm tokens, and 23 AWS keys, with 63 organizations among owners. In 539 of 1,105 traced repos, a commit listing Claude as co-author changed the file. One notified firm confirmed a live key; Apiiro did not test whether every secret still works. Remediation colour: rotate first, then git rm --cached, add a project .gitignore, and prefer pattern allows / env vars. Global git-excludes don’t travel with the project and don’t undo commits; Claude Code does not add the file to the project .gitignore. Of 16 coding agents compared, Claude Code is the only one confirmed to save the full approved command in a project file by default. Anthropic’s newsroom shows no new Nov 7–8 advisory slug (featured still Opus 5.5 Sep 22).

Watch next
  • Anthropic docs/changelog hardening for settings.local.json
  • Org-wide .gitignore templates for .claude/
  • Peer coding-agent audits of “don’t ask again” persistence
  • Whether Anthropic publishes a newsroom advisory
Limitation

THN reporting of Apiiro research (Apiiro first-party blog returned HTTP 500 this close — Wire did not independently read the blog body). Anthropic has not published a newsroom advisory. Live-exploit counts unknown. Vendor-research figures not independently reproduced by Wire. Distinct from Sat LIVE WorkOS ARTEX / FMF.

Cursor changelog (Nov 6, 2026) ships Voice Chat with Agents — two-way spoken calls in the Agents Window, IDE, web, and Cursor for iOS; the agent talks progress while coding; powered by Grok Voice (Ara, Eve, Leo, Rex, Sal); Privacy Mode → zero data retention. Soft title: platforms / coding-agent realtime voice, not a frontier model GA. Distinct from dictation (“Voice Mode”). Companion Nov 5 “Build Live Apps” changelog exists — not dual-carded here (desk soft-cut to D-notes). Sat-miss carry (not in Sat LIVE book).

Why it matters

Material coding-agent realtime voice surface across desktop/web/iOS with Grok Voice — weekend platforms Confirmed after Sat LIVE closed without it.

Qoder docs (promotion starts Nov 7, 2026 Singapore Time / UTC+8 = HKT) open a daily free allowance for stealth model Qoduck — soft title: models / coding-agent stealth access promo, not a named model GA and not open weights. Within allowance, Credit rate drops from 0.5× to 0.0× (no Credits, even at highest thinking intensity). Eligible: individual Free, Pro Trial, Lite, Pro, Pro+, and Ultra users of Qoder International (new and existing). Teams and Enterprise ineligible (do not see Qoduck in picker). Qoder IDE, JetBrains plugin, CLI, and other products currently not in the promo. End date TBD; Qoder may adjust allowance. After allowance: standard 0.5× Credit rate. Do not assert Alibaba/Qwen identity — page calls it a stealth model whose identity is not disclosed.

Why it matters

Same-window stealth coding-model free trial with zero-Credit-within-allowance economics — weekend models Confirmed without inventing a named frontier identity.

Ant Ling changelog (Nov 5, 2026) ships native multimodal Ling-3.2-flash — soft title: models / China multimodal MoE ship, not a US frontier GA and not an Inkling re-card. 560B total / 25B activated; 1M context; native image, text, and video input; focus: complex visual understanding, multimodal long-horizon Agent, 3D understanding & interaction, medical multimodal; chat UI plus OpenAI- and Anthropic-compatible APIs; free for the first 2 weeks after launch. Sat-miss carry. Soft: do not invent independent bench scores.

Why it matters

Material China multimodal MoE API ship with 1M context and dual-compat APIs — weekend models Confirmed distinct from Sat Inkling.

Anthropic Engineering (Nov 6, 2026) publishes “The components of agent memory” — soft title: platforms / agent-memory systems research, NOT a new Claude model GA and NOT on the /news product listing. Offline dreaming (sleep-time compute) corrects bad notes against transcripts, builds wiki-like memory maps, promotes notes into injected session context, and can focus memory on a goal. Evals: CL-Bench learning loops (later Claude models close the loop); Pokémon Red salience (Sonnet 3.7 wrote 413 memory files / 1 badge vs Fable 5 nine files + global.md / 4 badges in compared runs); Stainless corpus ~350,000 files / ~420M tokens — wiki built over 128 dreaming sessions, graded criteria 13% → 48%; wiki+search rose 54% → 61% as wiki grew vs ~53% search alone; wiki-only best 48%. Claude Code auto memory colour: loads only first 200 lines or 25 KB of MEMORY.md as injected header — optional depth colour, not a new product ship. Do not claim a user-facing “dreaming” switch shipped. Sat D-note elevate (not in Sat LIVE book).

Why it matters

First-party agent-memory systems research with measured dreaming evals — platforms depth without inventing a /news product GA.

More ▾ Less ▴
Confirmed
What happened

Anthropic Engineering on November 6, 2026 publishes “The components of agent memory.” Soft title: platforms / agent-memory systems research — NOT a new Claude model GA and NOT on the /news product listing. Stamp: Anthropic research/engineering Nov 6, 2026 (Sat D-note elevate; not in Sat LIVE book). Offline dreaming (sleep-time compute) corrects bad notes against transcripts, builds wiki-like memory maps, promotes notes into injected session context, and can focus memory on a goal. On CL-Bench, later Claude models close learning loops. On a Pokémon Red salience comparison, Sonnet 3.7 wrote 413 memory files and earned 1 badge versus Fable 5’s nine files plus global.md and 4 badges in the compared runs. On a Stainless corpus of about 350,000 files / ~420M tokens, a wiki built over 128 dreaming sessions lifted graded criteria from 13% to 48%; wiki+search rose from 54% to 61% as the wiki grew versus about 53% for search alone; wiki-only best was 48%. Claude Code auto-memory colour: the product loads only the first 200 lines or 25 KB of MEMORY.md as an injected header — optional depth colour, not a new product ship. Do not claim a user-facing “dreaming” switch shipped.

Watch next
  • Claude product surface that exposes dreaming controls
  • Peer lab memory papers
  • Claude Code auto-memory docs sync
  • Whether /news ever lists a dreaming product GA
Limitation

Research/engineering primary (not a /news product GA). Vendor-run evals, not independently reproduced by Wire. Product GA not claimed. Distinct from Anthropic IPO Developing chatter (no unlisted share prices; no S-1/A re-card).

Visa newsroom / Business Wire (Nov 5, 2026) ships the Visa Agentic Commerce Benchmark — soft title: open/reproducible / agentic-commerce eval, not a model ship. 22 models from 12 labs on 15 shopping scenarios; overall pass ~57.2%; fail 42.8%; about 54% of failed runs still ended in a purchase; all-three-attempts pass only 34.2%; best-performing models 82.2%, worst 28.9%. Public interactive experience on Visa Payments Frontier; plans to extend to checkout/payment/order. Trust Index colour (72% used AI assistant / 23% trust AI for payments — Sep 2026) is context, not the card stake. Do not invent named model leaderboard identities (PR does not fully list them). Sat-miss carry.

Why it matters

First-party agentic-commerce eval with published pass/fail rates — open/reproducible weekend Confirmed without inventing a model leaderboard.