LLMs corrode institutional failure-learning; Rocket Lab acquires Iridium's satellite constellation; one V8 compiler bug escapes Chrome's renderer sandbox

Nebusec

Longinus: Single V8 Compiler Bug Pierces Chrome's Renderer and Sandbox

CVE-2026-6307 is a V8 flaw in which the FrameStateFunctionInfo equality operator fails to compare WebAssembly function signatures, allowing common subexpression elimination to incorrectly merge deoptimisation states for JS-to-Wasm calls with different return types (i64 vs externref). The exploit chain uses this to materialise an externref as an integer (an addrof primitive) and an i64 as a fake object reference (a fakeobj primitive), then leverages the forged object to escape the V8 heap sandbox and reach arbitrary memory — achieving full renderer RCE with no additional vulnerabilities required. The architectural significance is that a single compiler correctness bug in the inlining layer simultaneously crosses two sandboxing boundaries, demonstrating that the layered V8/Chrome security model can be defeated at the optimisation stage rather than at the more heavily-audited memory-management layer. The writeup includes a complete exploit chain with reliability-focused techniques for seeding JIT-compiled code at predictable addresses.

SpaceNews

Rocket Lab Acquires Iridium for $8 Billion

Rocket Lab announced an $8 billion acquisition of Iridium Communications ($54 per share, a 24% premium) in a deal expected to close mid-2027. Rocket Lab brings launch capability and manufacturing; Iridium contributes an operational 66-satellite LEO constellation with rare L-band spectrum, $871.7M in annual revenue, and established global communications and aviation-tracking services that Rocket Lab would otherwise take a decade to build. The combination creates a vertically integrated space infrastructure company — from launch to end-user connectivity — at a moment when the industry is consolidating around operators who control both the delivery and the service layer. For the space sector, the deal signals that independent launch providers see a commodity-risk ceiling in launch-only business models and are moving to capture margin further up the value chain.

GitHub Blog

GitHub Advisory Database: Record Output, Unprecedented Backlog

In May 2026, the GitHub Advisory Database published 1,560 reviewed advisories — five times its typical monthly output and a new record — while still failing to keep pace: private vulnerability reports surged from ~550 to ~3,000 per week, repository advisories from ~650 to ~5,000 per week, and GitHub CNA CVE requests hit nearly 4,000 in May alone. Processing times extended to multiple weeks for a meaningful share of advisories, because the inbound surge brought proportionally more complex cases requiring package disambiguation across ecosystems, version range reconstruction from commit history, and resolution of conflicting upstream data — work that resists parallelisation and doesn't compress with volume. GitHub frames the surge as structural, not cyclical: 30,000+ CVEs have already been published in 2026 and 1.7 million repositories now have private vulnerability reporting enabled, meaning the ecosystem has crossed a threshold where disclosure volume is itself a constraint on security response infrastructure. The team is deploying AI-assisted research tooling while holding the line on human validation for every reviewed advisory — a reasonable trade-off that makes the throughput bottleneck visible as a systems design problem, not just a staffing one.

Fergus Finn

What Happens When You Run a CUDA Kernel

When a CUDA kernel launches, the compiler inserts a stub that packs arguments and calls the usermode driver (libcuda.so.1), which queues commands in a pushbuffer ring structure coordinated by a GPFIFO register the GPU polls to track consumed work. The GPU's host engine reads a queue metadata descriptor and dispatches thread blocks across SMs, where four independent schedulers manage up to 48 resident warps using compiler-generated control codes embedded in each 128-bit instruction — stall counts and scoreboard barrier indices allow warps blocked on memory to be skipped without OS involvement. Memory accesses coalesce automatically: 32 per-thread 4-byte loads become four 32-byte sector requests, a hardware behaviour that makes the access pattern the primary determinant of effective bandwidth. The article is a thorough systems-level walkthrough from system call through MMIO doorbells to warp scheduling, useful for diagnosing whether a kernel is memory-bound, register-pressure-bound, or occupancy-limited.

htmx.org

Working With AI: A Concrete Example

Carson Gross describes debugging a hyperscript parser bug with Claude: the AI correctly diagnosed the root cause but proposed two inadequate fixes — one too narrow, one architecturally overbroad — before Gross identified the right solution using hyperscript's 'follows' mechanism, a domain-specific construct the AI had encountered but didn't weight correctly. The diagnostic insight is that AI excels at investigation (reproducing bugs, tracing execution paths) and test generation but lacks the architectural taste to distinguish between a fix that works and a fix that's right — a distinction engineers carry as tacit knowledge. Gross frames AI coding assistance as a tool that performs best when a skilled engineer remains in active dialogue rather than accepting outputs passively, and notes it compensates well for age-related memory limitations while preserving design judgment. The essay is an honest counterpoint to both dismissive scepticism and uncritical enthusiasm about AI pair programming.