Top stories

  1. Cloudflare discloses cross-request KV cache exposure in GLM-5.3 on Workers AI caused by an SGLang bug An SGLang bug on Workers AI let 155 GLM-5.3 responses include text from other customers' requests between 2026-08-27 and 2026-09-09.
  2. Warp says attackers used a vendor flaw to reach agent trace and telemetry data for up to 1% of users Warp's CEO says attackers exploited a vendor in its infrastructure and reached agent trace and telemetry data for up to about 1% of users.

Security

  1. RustSec publishes Pingora HTTP/2 and header advisories and a batch of zstd-safe soundness bugs RustSec added HIGH advisories for Pingora HTTP/2 memory exhaustion and Connection header ACL bypass, plus over a dozen zstd-safe soundness bugs.
  2. Grafana ships 13.3.0 and patch releases back to 12.4.14 fixing multiple CVEs Grafana's 2026-10-27 releases patch up to ten CVEs in 13.3.0, and every supported branch down to 12.4 gets at least three of those fixes.
  3. Wiz sees PHP2Shell exploit attempts fail in its telemetry as vulnerable exposure falls about 67 percent Wiz honeypots caught public PHP2Shell kits exhausting PHP memory but failing to leak addresses, while exposed vulnerable hosts dropped sharply.
  4. Go security team details its first URGENT release distrusting the weak eMudhra intermediates Go's security team explains how Go 1.27.3 and 1.26.10 shipped in under 7.5 hours to distrust the weak eMudhra intermediates.
  5. FoxIO details three GoAhead TLS bugs, including ignored CRLs that accept revoked client certificates The embedded GoAhead web server loaded CRLs without enforcing them and truncated a DH modulus to one byte, both fixed in GoAhead 6.1.1.

AI

  1. OpenAI releases GPT-6.1 Terra, a midrange model for visual reasoning and computer use OpenAI's new midrange model targets visual reasoning and computer use at $1 per million input tokens and $5 per million output tokens.
  2. Debian developer stops NEW uploads after the queue starts adding unannounced LLM reviews Debian's NEW queue now attaches an LLM-generated review to uploads, and one developer has stopped using NEW until an opt-out exists.

ML research

  1. OpenAI publishes a model-generated claim that the Exponential Time Hypothesis is false OpenAI published a model-generated argument and Lean formalization claiming the Exponential Time Hypothesis is false, pending outside review.

Developer tools

  1. Microsoft documents the Type Server Protocol for querying Python type analysis over JSON-RPC Microsoft documents a JSON-RPC protocol that lets tools query a Python type server for computed, declared and expected types per snapshot.
  2. Ghostty 1.4.0 ships with ten security advisories, full Kitty graphics support and scrollback compression Ghostty 1.4 fixes ten advisories, including OSC 8 link command execution on macOS, and adds full Kitty graphics and scrollback compression.
  3. JetBrains IDEs 2026.3 replace the JediTerm terminal emulator with libghostty JetBrains IDEs 2026.3 replace the JediTerm terminal emulator with libghostty, adding Kitty keyboard protocol and theme change notifications.

Languages and runtimes

  1. CPython's experimental JIT faces removal from main after its authors withdraw PEP 836 The authors of PEP 836 withdrew it, so CPython's experimental JIT from 3.13 is expected to leave the main branch early in the 3.16 cycle.
  2. OpenJDK asks projects to test JDK 28 with previews on, as Integer and java.time types become value classes With preview features enabled, 30 JDK classes lose identity, so == can return true, synchronizing on them fails, and WeakHashMap keys throw.

Linux and kernel

  1. Linux 7.4 merges DMA-BUF read and write support for NVMe through io_uring registered buffers A dmabuf can now be registered with io_uring and used with READ_FIXED and WRITE_FIXED, enabling zero-copy device memory I/O to raw NVMe devices.

Engineering posts

  1. Meta describes PrivacyLib, which enforces privacy rules inside data system clients, including agent tool calls Meta describes PrivacyLib, a library inside clients of hundreds of data systems that runs billions of privacy checks per second.

Markets and companies

  1. Anthropic files S-1 disclosing up to $120.5B in SpaceX GPU compute commitments through 2029 Anthropic's IPO filing commits up to $120.5B for SpaceX GPU compute through 2029, cancelable on 90 days notice after early 2027.