Today’s lead · TheHackerNews
Claude Code’s “don’t ask again” can leak secrets into Git
A Claude Code convenience feature has become a secrets-management footgun: Apiiro says 1,122 public .claude/settings.local.json copies contain passwords or keys saved from approved shell commands. If your team uses “don’t ask again,” rotate exposed secrets before cleanup; git history and forks keep leaks alive.

Top signals
4 moreTools & repos
5 selected
Exa Monitors v2
Exa Monitors v2 gives agents scheduled web search with webhook delivery only when something new and relevant appears. The practical pitch: fewer polling loops, citations, JSON output, and a lower $2 per 1k runs price.
cloudflare/security-audit-skill
Cloudflare’s security-audit-skill packages a multi-phase security audit workflow for coding agents, with independently verified, machine-readable findings. Useful if your agent reviews need structure instead of another vague “look for bugs” prompt.
vercel-labs/agent-browser
agent-browser is Vercel Labs’ Rust CLI for browser automation by AI agents. The repo is a clean signal that browser control is becoming agent infrastructure, not just a demo layer.

LiveReview
LiveReview pitches risk-aware PR review: auto-approve low-risk changes, route risky ones to the right engineers, and reduce senior reviewer load. Treat the production-impact claims as vendor claims until your own incident data agrees.
google/skills
google/skills is a Python repository of agent skills for Google products and technologies. It is worth watching as large vendors standardize reusable action packs for agents.
Blogs worth your time
1 reads
Two Minute Papers cuts through the Astra weight-editing hype
The video explains MIT research where GPT-6 Astra reads, edits, and sometimes improves neural-network weights, while pushing back on overclaims. The sober takeaway: narrow, custom experiments are impressive enough without pretending models can generally rewrite each other.
Community discussions
3 threads
Agentic dev may make CI runners the next cloud bill fight
The argument: once agents spray PRs, runner and sandbox costs can dwarf everything except tokens. The author says productized CI runners carry a huge premium over DIY orchestration on cloud or owned hardware, especially for single-tenant setups.
Local agents on 16GB VRAM: useful, but setup-heavy
A 16GB VRAM user asks why local helper, research, and dev agents still feel underwhelming. Replies are pragmatic: small setups can work, but model choice, quantization, harnesses, and task scope matter more than the headline VRAM number.
A Claude thread reminds builders that benchmarks are not daily-driver tests
A GPT user says Opus 5.5 felt dramatically better on real R&D workflows than benchmark gaps suggest. The useful tension in the replies: subjective fit is real, but limits, domain variance, and repeatable task-based comparisons still matter.
Funding & acquisitions
1 movesScuba emerges with $34M for shared team memory in the agent era
Scuba says it is coming out of stealth with $34M to build shared context for teams and agents across Slack, Claude, ChatGPT, Cursor, and MCP tools. The bet: agent productivity bottlenecks move from model intelligence to organizational memory.
Bengaluru radar
2 events
Claude Builder Day Bengaluru: Maximizing Margins
A hands-on Anthropic × AIBoomi build session in Indiranagar for startup CTOs and engineering leads optimizing AI-native SDLC cost.
Great International AI Native Summit (GAINS)
A Bangalore in-person conference listing for Dec. 9 aimed at engineering teams working across AI, cloud, data, DevOps and architecture.

