2026-11-14
Top stories
- Unit 42 says China-nexus attackers used the open-source Cairn agent to attack 19 organizations with little human input A customized Cairn pentest agent on GLM 5.2 and DeepSeek V4 ran 25 operations alone and got access or data at 10 targets, Unit 42 says.
- Cursor traces its Claude Code extension backdoor to a Fireworks-exposed key and agent traces holding credentials Attackers copied Cursor agent traces from a bucket shared with Fireworks, then used keys in them to backdoor an extension for up to 5,000 users.
- Project Zero drops its 90+30 policy and will disclose bugs 90 days after report, falling to 30 days by 2028 Google Project Zero now discloses bugs 90 days after reporting, fixed or not, then 60 days from February 2027 and 30 days from January 2028.
Security
- Mior and Skyfall let a malicious host read AMD SEV-SNP guest memory, with BIOS fixes due November and December Two attacks break AMD confidential VMs on EPYC chips, one by corrupting the RMP table at boot and one by leaking leftover VM state via TSA.
- CISA adds actively exploited UnrealIRCd CVE-2026-114713 to KEV, a label tag flaw that injects server commands Any user on a network of linked UnrealIRCd 6 servers can inject server commands via an unescaped label tag, and attackers now do.
AI
Agentic coding
- Claude Code 2.1.325 makes sandboxed Bash commands ask for permission by default Sandboxed shell commands no longer run without prompts unless sandbox.autoAllowBashIfSandboxed is set, and scheduled skills now ask first.
- Coral cuts agent cost per turn about 65% by compacting idle agents past 225,000 tokens of context Cache reads of 430,000 to 550,000 tokens per turn drove the bill, and compacting idle agents dropped average context to 178,000 tokens.
- Google's Antigravity /teamwork agents finish a Rust rewrite of Zircon, adding 800,000 lines that boot on arm64 Google Research says its multi-agent harness ported the rest of Zircon and the libgav1 AV1 decoder to Rust, checked against unmodified tests.
- nono 0.82 adds session terminal recording, live proxy audit logs and network approvals for plain-HTTP requests The nono agent sandbox now records terminal sessions, logs proxy audit events live, and asks for approval on plain-HTTP and Unix socket access.
Developer tools
- Ratatui maintainer Josh McKinney proposes a cell-border terminal protocol, with a working Ghostty patch A DCS sequence lets each cell carry up to twelve colored line segments, so TUI panes get borders without losing a row or column.
- PyPI adds a security tab with per-file attestations and warnings when a release's provenance changes or disappears PyPI project pages now show whether each release file has an attestation and warn when provenance is incomplete, inconsistent, changed or lost.
Languages and runtimes
- Luau 0.747 adds lazy bytecode loading, buffer vector reads and writes, and native codegen userdata access The new luau_loadlazy builds only the main function at load, leaving other functions as lazy stubs loaded on first use.
- The ant Go ORM uses Go 1.27 generic methods to check column names at compile time without codegen Pointers to fields of a pinned probe struct stand in for column names, so a misspelled column or wrong value type fails go build.
- Primitive patterns will be withdrawn instead of returning in JDK 28, while ZGC gains adaptive heap sizing Oracle's Inside Java Newscast says value-class numeric conversion questions stall primitive patterns, as JEP 546 tunes the ZGC heap on its own.
Linux and kernel
Engineering posts
- pg_orca prices ORCA plans with a port of PostgreSQL's costsize.c, matching stock planner costs within 2% pg_orca ports PostgreSQL's costsize.c into the ORCA optimizer so its plan costs match the stock planner, checked by a harness within 2%.
- Unread no-store fetch replies leak one file descriptor each in Chromium, freezing Claude Desktop on Linux A developer traced Claude Desktop hangs to Chromium keeping a 2 MiB shared-memory pipe and fd open for every unread no-store response body.
Videos
Hacker News
- Jonathan Blow interview questioning AI coding productivity draws 261 comments on Hacker News Blow argues LLMs act as fast search but fail to track constraints across large programs, and says he sees no visible jump in software output.
- FAA's 4.5-hour nationwide ground stop draws 1,567 points and 1,188 comments on Hacker News Practitioners debate Friday afternoon deploys and circuit versus packet switching after a reported network update halted all US departures.