2026-10-29
They compiled a working RISC-V CPU into an LLM’s weights
The AI didn’t call a calculator. It *was* the calculator, compiled into its own neurons.
A developer compiled a 32-bit RISC-V CPU (RV32IM) directly into the weights of a stock Qwen 3.5 4B model, using the model’s own Gated DeltaNet layers as exact, addressable RAM. It executes C and MicroPython inside forward passes with no external tools or sandbox, passed official RISC-V tests, matched QEMU, and ran 18 million instructions error-free. Arithmetic benchmarks jump from ~55-72% (base model) to 97.5%+, matching or beating a Python interpreter tool.
What Happened
Divyansh Agrawal’s technical write-up, GitHub/Hugging Face checkpoints, Hacker News (553 pts).
2026-10-29
A fingerprinting script packed with bioweapon function names to stop AIs from reading it
They named the functions after ricin so the AI would refuse to look at the code.
While crawling 10,000 sites, a researcher found a Russian anti-DDoS company’s browser fingerprinting script whose function names include computeRicinLethalDosageForWaterSupply and an encoded Unicode jailbreak prompt demanding unrestricted guides to bioweapons, DDoS on hospitals, and tax evasion. The names trip safety filters so AIs refuse to deobfuscate it; a Chinese model still summarized it. The actual script is ordinary canvas/anti-bot telemetry.
What Happened
nullpt.rs reverse-engineering post, Hacker News (1,341 pts), X.
2026-10-29
A reporter gave the AI Death Clock his blood. It started sucking the life out of him
I gave an AI my blood so it could tell me when I die. Then it started selling me more blood tests.
Death Clock, a longevity app, predicted the writer would die August 10, 2060 at age 74 (cartoon grim reaper on a donut float). He paid for a $69 blood panel plus add-ons; the date moved, notifications piled up, and after a month he felt the app was draining him rather than extending life. The AI is trained on 1,200 studies plus insurer tables and upsells more tests.
What Happened
WIRED first-person, Oct 29.
2026-10-29
Claude will now tell your boss when your chats were “off-topic
Your company paid Claude to read your chats and tell the boss which ones were off-topic.
Anthropic shipped “smart reports” for Claude Enterprise. Managers pick a team and date range; Claude reads a sample of transcripts and graphs workstreams, cost, deliverables, frictions, and “inefficiencies”—including sessions classified as personal or off-topic (shown only as aggregates). Anthropic says it is not for individual performance reviews. Delegated team leads can run reports on their groups.
What Happened
Mashable, Anthropic help docs and product post, Oct 28–29.
2026-10-29
An AI agent found four zero-days in the software that runs particle accelerators and fusion labs — in 3 hours 49 minutes
The AI that found the bugs in the software that runs fusion reactors did it before lunch.
Cipher, an autonomous agent, discovered four memory-safety bugs (one CVSS 9.8 potential RCE) in EPICS Base, the decades-old control toolkit used at particle accelerators, telescopes, and ITER’s CODAC fusion systems. No prior EPICS knowledge, 3 h 49 m, coordinated disclosure, patched in 7.0.11. The bugs had sat 14–31 years.
What Happened
Causal Security write-up, CISA advisory ICSA-26-300-01, SecurityWeek, Oct 27–28.
2026-10-29
Google’s new #1 model is so good at finding and exploiting bugs they will only give it to cyber defenders for now
Google built an AI that’s #1 at finding and exploiting vulnerabilities — so they’re not letting you use it yet.
Gemini 4 Carbon, announced two weeks after Argon, ranks #1 on Artificial Analysis v5.0 (59). It leads SEC-bench Pro (88.2% finding vulns), ExploitGym (54.8% exploitation), and CWE-bench (71% patching). Google is rolling it first to a subset of cyber defenders via the Fairwind Program and U.S. pre-release testing; developers and the public wait while guardrails are tested. Same intro price as Argon.
What Happened
Google DeepMind blog, Ars Technica (“the frontier AI model that you can’t use”), The Verge, Oct 28.