Claude finds two critical Xen VM-escape flaws; patch 4.6 and up

XSA-523 lets a hostile guest read or write another VM's memory, dom0 included. Also: agent-cost papers, Qualcomm's 128GB card, and agent payments over MCP.

Nowline NOV 5 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Xen XSA-523: one VM can read and write every other

    Two grant-copy flaws (CVE-2026-102703 and CVE-2026-102704) let a malicious guest read and write other VMs' memory, including dom0 — found by Anthropic's Claude and confirmed by Ada Logics. Every Xen release from 4.6 up is affected; patches cover 4.17.x–4.19.x and xen-unstable, but 4.6–4.16 are end-of-life with no fix. If you run Xen, patch or migrate today.

  • Paper: skip long context, cut agent costs up to 54x

    A new paper argues long context is mostly wasted in agent loops and proposes End-of-Conversation and Single-Action Compaction, reporting 6.5–54.6x cost savings while accuracy shifts just -3.7 to +5.6 points. If your agent replays full history every step, compaction is a weekend-sized refactor with real token savings.

  • Paper: your draft model quietly steers long agent runs

    Researchers (MIT, Princeton, Northeastern) report that speculative-decoding rejections shift an open-weight model's output after roughly 30 steps in long-horizon agent tasks — and that tuning this lifts accuracy without touching the main model's weights. A reminder that your draft/verifier pairing is a behavior knob, not just a speed trick.

  • Qualcomm's Cloud AI 100 Ultra: 128GB inference at 150W

    Qualcomm listed a single-slot PCIe inference card carrying 128GB of LPDDR4x at 548GB/s while drawing just 150W — enough memory to hold a large model on one card without splitting it across GPUs. Pricing and availability aren't disclosed yet, so treat it as a spec sheet to watch rather than something to buy today.

  • Pliant opens card issuing and payments to agents via MCP

    Berlin's Pliant raised an $80M Series C and now exposes card issuing and real payments to agents like Claude and ChatGPT over MCP. If you're building an agent that needs to actually buy things, there's a programmable-card backend that speaks your protocol — worth a look before you roll your own.