Tuesday, July 7, 2026
An AI agent ran a full ransomware intrusion, Meta covertly probed rivals' chatbots with fake teen accounts, and Nvidia's next-generation rack slipped to 2028.
An AI agent ran the entire ransomware intrusion
Sysdig documents what it describes as the first ransomware intrusion run end to end by an autonomous LLM agent, attributed to a group it names JadePuffer. The claim is single-vendor and not independently verified. Per Sysdig, the agent gained access through CVE-2025-3248, an unauthenticated RCE in the Langflow LLM-app framework patched on April 1, 2025, then ran reconnaissance, credential theft, lateral movement to Alibaba Nacos, privilege escalation via CVE-2021-29441, and encryption with no human in the loop. It corrected a failed login in 31 seconds and switched its payload parsing when a server returned XML instead of JSON. Execution was flawed: it encrypted 1,342 Nacos config items with MySQL’s AES_ENCRYPT(), but Sysdig assesses the cipher as AES-128-ECB, not the claimed AES-256, and the randomly generated key was never stored or transmitted, so decryption may be impossible regardless of payment. Two byproducts double as detection signals: verbose natural-language comments in the attack code, and a ransom note carrying a placeholder Bitcoin address apparently reproduced from training data.
Safety framing, and silent rewrites
WIRED reports Meta ran a covert project, codenamed Cannes and run through contractor Covalen, in which hundreds of workers created fake under-18 accounts on ChatGPT, Gemini, and Character.AI and sent high-risk prompts, framed internally as safety benchmarking. One August 2025 round logged more than 45,000 prompts; WIRED reviewed a separate 3,748-prompt spreadsheet with hundreds targeting suicide and self-harm and at least 239 involving sex or romance. None of the three companies authorized the testing, and all said it violated or may violate their terms. Meta calls the work responsible, industry-standard practice and says it does not train on competitor outputs; the reviewed documents do not show how Meta used the responses. Humane Intelligence’s Rumman Chowdhury called the scale, secrecy, and minor-impersonation method outside normal evaluation, and a possible case of safety serving as cover for competitive intelligence.
Separately, a study from Oxford’s Internet Institute and the Hasso Plattner Institute found LLMs from xAI, Meta, Google, Alibaba, and Mistral introduced political bias when improving user drafts, even when told to preserve meaning. Four skewed liberal; Grok skewed conservative, which the researchers attribute to X’s instruction to challenge mainstream narratives. Mistral turned a “#climatechangehoax” post into a climate-action appeal; Qwen rewrote “Jesus wasn’t real” into an affirmation. The authors say neither the EU AI Act nor the Digital Services Act covers AI-mediated rewriting; the Guardian write-up does not link the paper or give sample sizes.
Datacenter hardware
SemiAnalysis reports Nvidia’s Kyber NVL144 rack, built to hold 144 Vera Rubin Ultra GPUs, has slipped more than 12 months to 2028 over the manufacturability of its PCB midplane. The eight-rack NVL576 is also said to be delayed or held to low volumes. A stopgap pairing two current racks was cancelled after hyperscaler objections, leaving, in SemiAnalysis’s words, no proven way to expand the scale-up world size for Rubin Ultra. Current Rubin systems ship on schedule this fall; Nvidia did not comment.
Huawei will enter South Korea’s accelerator market in Q4 2026 with the Ascend 950PR inference chip, claiming about 2.87x the inference throughput of Nvidia’s H20 at roughly a quarter the cost. Both figures are self-reported, and the H20 is Nvidia’s export-restricted China part, not its flagship. Huawei’s scale bet is the Atlas 950 SuperPod at 8,192 accelerators per deployment, with distribution already signed through Hansol PNS and SK Shieldus.
Inference infrastructure, measured
Moondream’s Photon engine overlaps CPU housekeeping with GPU decode to hide “GPU bubbles,” the idle gaps between autoregressive steps. On moondream2 it reports speedups from 6.5% (RTX 3090, one stream) to 35.4% (B200, 32 streams), with an analytic model matching within about four points; the bubble is a fixed CPU cost, so it grows as GPUs speed up or models shrink. Databricks details its production GPU-reliability stack and a subtle hazard: NCCL’s InfiniBand timeout defaults to about 7 seconds, far below PyTorch’s roughly 10-minute watchdog, so one long port-down event can kill a multi-day run before the watchdog fires. At a stated 1% annual failure rate, a 1,024-GPU, 30-day job carries a 57% chance of at least one failure.
What to watch today
- Whether a second security vendor corroborates Sysdig’s claim that JadePuffer ran without human input; the milestone rests on one report.
- OpenAI said it is “looking into” Meta’s testing; watch for a terms-of-service enforcement move against Meta or Covalen.
- Huawei’s Ascend 950PR reaches South Korean commercialization in Q4 2026 through Hansol PNS and SK Shieldus.
- Whether Nvidia confirms or disputes SemiAnalysis’s 2028 Kyber slip.