GPT-6.2 Astra crosses OpenAI’s “Critical” cyber line

By Mark 11 min read 0 views

😁 Hello, super humans! Model launches usually come with a price cut and a pile of benchmark charts, and this one has both. The part worth slowing down for is the safety paperwork: OpenAI’s own system card says its newest models crossed the top rung on cyber capability. And as if to prove the point, a security team used an AI coding agent to find a critical bug in Java in about an hour. Let’s dig in.

📰 Quick Signals

  • 🧠 AI: Anthropic reportedly priced its IPO at $83 a share, above its $70 to $78 range, for a fully diluted valuation of about $1.5T; trading on Nasdaq under “ANT” is expected to start today.
  • 🤖 Robotics: GlobalFoundries will be the first non-automotive customer to test Boston Dynamics’ Atlas humanoid, onsite from early 2027, after two Spot robots in its Singapore fab logged about 900 km and more than 90,000 sub-fab inspections since April at a 99% mission success rate.
  • 💻 Programming: “Java4Shell” (CVE-2026-114938) is a memory-corruption bug in HotSpot’s AVX-512 Base64 intrinsic that can lead to pre-auth RCE in apps like Keycloak and Jenkins; Oracle and OpenJDK shipped fixes on Nov 17, and Calif says an AI coding agent helped find it in about an hour.
  • ⚡ Electronics: JEDEC previewed DDR6: 74 mm modules instead of DDR5’s 133 mm, a 12.8 GT/s baseline (14.4 GT/s peak), roughly 4x DDR5’s baseline bandwidth per module and up to 1 TB per slot; no launch date yet.
  • 📡 Telecom: T-Mobile told the FCC it is buying Bluewater Wireless’ 600MHz licenses across 435 counties in 27 states, adding about 10MHz in most areas and keeping that low-band spectrum out of SpaceX’s reach.

🔍 The Big Story: GPT-6.2 Astra crosses OpenAI’s “Critical” cyber line

If you build agents on OpenAI’s API, you just got a cheaper flagship; if you defend systems, you just got a vendor admitting its models are now very good at attacking them.

What happened: OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on Nov 17. Astra is the frontier model for the hardest projects; Sol is the efficient everyday model for multi-step work. Both are live in ChatGPT, Codex and the API as gpt-6.2-astra and gpt-6.2-sol. The system card rates both models Critical for cybersecurity and High for biology and chemistry, and makes Astra the first externally released OpenAI model rated High for AI self-improvement.

The details: Pricing moves in the right direction. Astra costs $8 per million input tokens, $0.40 cached and $40 output, 20% below GPT-6 Astra, while Sol stays at $2 / $0.10 / $10. OpenAI claims Astra beats GPT-6 Astra by 5.2 points on Agents’ Last Exam at under a third of the cost per task, and gains 6.2 points on its internal GDPval v2 while using 22% fewer output tokens; Sol lands within half a point of Astra on Cognition’s FrontierCode 2.0 at about a fifth of the cost. These are vendor numbers, and early hands-on reviews such as Simon Willison’s are mixed. The system card is the more interesting read. On cyber evals Astra and Sol score 86.75% and 88.34% on SEC-Bench Pro, and both hit 100% on ExploitBench V8. In response OpenAI tightened default limits on higher-risk cyber help, requires extra vetting for permissive access, and monitors internal Codex traffic, RL runs and external use. For self-improvement it gates internal expansion on alignment evals, isolated research environments more tightly, and added a “Control Room” system that tracks research workloads and can halt them. It also notes Astra showed more evaluation awareness than GPT-6 Astra, flagged in under 3% of turns.

flowchart TD
    A[GPT-6.2 Astra and Sol<br/>launched Nov 17] --> B[Preparedness Framework v2 ratings]
    B --> C[Cybersecurity: Critical<br/>both models]
    B --> D[Bio and chem: High<br/>both models]
    B --> E[AI self-improvement: High<br/>Astra only, a first]
    C --> F[Tighter default cyber limits,<br/>vetting for permissive access,<br/>monitoring of internal and external use]
    E --> G[Alignment evals gate internal use,<br/>isolated research environments,<br/>Control Room can halt workloads]
    C -.real-world example.-> H[Java4Shell: AI agent finds<br/>critical Java RCE in about an hour]

Important

Our take: I would not pick a model off these benchmark charts; run your own evals, because Sol at a fifth of Astra’s price is probably the right default for most agent loops. What I would act on is the cyber rating. When the vendor itself says the model is at its top cyber tier, and the same week a researcher turns an AI-assisted audit into a critical Java RCE for under $200 in tokens, the patch window you used to rely on is shrinking. Treat “patch within days” as the new baseline, and start pointing these models at your own code before someone else does.

🗞️ More News

🧠 AI

  • Microsoft launched MAI Code 2 and MAI Code 2 Flash; it claims Code 2 reaches 98% of GPT-6.1 Sol’s performance at 40% of the cost (at promotional pricing), with Code 2 in GitHub Copilot and Flash in limited preview in Foundry.
  • Cerebras now splits inference across vendors: AMD Instinct GPUs do prefill and CS-3 systems do decode, giving 5x throughput per megawatt on a GPT-OSS-120B endpoint versus CS-3 alone, with AWS and Fireworks AI in early access.
  • Citi updated its online agreement so customers who give third-party AI agents account access are responsible for that activity, even if the tool is hacked; JPMorgan, Bank of America and PNC have made similar changes.
  • Nvidia says it will stop selling H200 chips in China after Jensen Huang’s ouster and new US rules on foreign data centers serving Chinese customers; Jefferies estimates the rules could cut $44B to $62B from next year’s data center revenue.
  • In GitHub and Microsoft’s MDASH pilot, an agentic scanner flagged 9,791 possible issues in six open-source projects; after triage only 99 private reports with proposed fixes went to maintainers, and 55 were fixed by Nov 15.
  • Google added temporary threads to AI Mode in Search, conversations that stay out of search history and personalization, rolling out globally on desktop and mobile.
  • OpenAI’s Astra for Law, built on GPT-6.2 Astra, now passes every required check on 53.8% of legal research questions, up from about 46% in September; enterprise access is on request.

🤖 Robotics

  • Bengaluru’s Humyn Labs raised $47M ($26M equity led by Accel plus $21M venture debt) to supply training data to US robotics developers; it reports $10M revenue from April to October and $20M in signed contracts.
  • Einride’s contracted ARR grew 76% since June to $167M, with driverless hours rising from about 5,400 to 8,500 and more than 40 customers; it targets cash-flow breakeven in 2028.
  • Kawasaki’s F Series S-type controller ships with CoreCubic-S Lite safety software (assessed to ISO 10218-1:2025) and built-in cybersecurity aimed at the EU Machinery Regulation and Cyber Resilience Act deadlines.
  • Penguin Random House’s Barcelona warehouse scaled from 14 to 40 Locus robots and now picks more than 510 units per hour, up from 253 manually.

💻 Programming

  • Malicious agent-framework releases 1.26.1 and 1.26.2 hit PyPI 33 minutes apart, running a credential stealer from a .pth file on every Python start; pin 1.26.0 and rotate secrets on any machine that installed them.
  • Solid 2.0 is stable after five years: async is part of the reactive graph, server functions get persistent live streams, several APIs like createResource and batch are gone, and SolidStart is retired.
  • jQuery 4.1.0 adds jqXHR.responseURL and a ReDoS fix, while 3.8.0 is the last planned 3.x feature release; from here 3.x only gets critical fixes.
  • GitHub detailed the Agent Host Protocol, an open spec that lets a Copilot agent session continue across VS Code, the Copilot app, mobile and cloud, with client libraries in six languages.

⚡ Electronics

  • Bluetooth 7.0 is ratified with optional High Data Throughput up to 7.5 Mbps on a new LE PHY; in a Nordic demo a 2 MB file moved in about 4 s versus 12 s on LE 2M.
  • de:volt is a free browser breadboard simulator with 105 parts, Arduino, micro:bit and RP2040 boards, MicroPython, and a built-in 4-channel oscilloscope and multimeter.
  • A Bank of America report says consumer NAND spot prices dropped 20%, which could make SSDs, laptops and phones cheaper.
  • Three decommissioned Nvidia V100s in a $1,042 homelab rig hit about 100 tokens/s on Qwen 3.6 35B, outpacing a far pricier Mac Studio at long contexts.

📡 Telecom

  • VodafoneThree will lab-test, then field-trial next year, Nokia’s AI-native anyRAN software on Nvidia’s Aerial platform, benchmarking capacity and spectral efficiency against its commercial network.
  • AMD’s telco lead predicts a “backlash” against power-hungry GPUs in the RAN, pitching 84-core CPUs that run forward error correction in software instead.
  • Juniper Research forecasts just over 47,000 private LTE and 5G networks in service by 2031, up from about 10,000 in 2026, with 5G reaching 52% of them.
  • Tucows agreed to sell Ting’s owned fiber markets, about 170,000 homes and businesses in seven markets, to Clearwave, removing about $301.5M of securitized debt at closing in Q1 2027.
  • Quad9 is withdrawing its free 9.9.9.9 resolver from France after court-ordered daily penalties for not blocking domains, and Cloudflare warns DNS-level blocking orders are spreading.

👨‍💻 Code Corner

Java4Shell only bites on x86-64 CPUs with AVX-512 VBMI running an unpatched JDK, so the first triage question per host is “do I have both?” This script answers it on Linux using the CPU flags Calif lists and the fixed versions from Oracle’s Nov 17 update.

import re
import subprocess

REQUIRED_FLAGS = {"avx512f", "avx512bw", "avx512vbmi", "bmi2"}
# First fixed Oracle builds per feature release (Nov 17, 2026 update).
FIXED = {17: (17, 0, 21, 1), 21: (21, 0, 13, 1), 25: (25, 0, 5, 1), 27: (27, 0, 1, 1)}

def cpu_has_vbmi_path() -> bool:
    try:
        with open("/proc/cpuinfo") as f:
            for line in f:
                if line.startswith("flags"):
                    return REQUIRED_FLAGS <= set(line.split(":", 1)[1].split())
    except FileNotFoundError:
        pass
    return False

def java_version() -> tuple[int, ...] | None:
    try:
        out = subprocess.run(["java", "-version"], capture_output=True, text=True).stderr
    except FileNotFoundError:
        return None
    m = re.search(r'version "([\d.]+)', out)
    return tuple(int(p) for p in m.group(1).split(".")) if m else None

cpu = cpu_has_vbmi_path()
ver = java_version()
print("CPU has AVX-512 VBMI path:", cpu)
print("Java version:", ".".join(map(str, ver)) if ver else "not found")

if ver and ver[0] in FIXED:
    patched = ver >= FIXED[ver[0]]
    if patched:
        print("OK: at or above the first fixed build")
    elif cpu:
        print("AT RISK: unpatched JDK on an affected CPU, patch now "
              "or start the JVM with -XX:-UseBASE64Intrinsics")
    else:
        print("Unpatched, but this CPU does not use the vulnerable path; patch anyway")
elif ver:
    print("Feature release not in the fixed list (8 and 11 are reported unaffected)")

Run it across a fleet with your config tool of choice and you get a quick map of which boxes need the patch first and which can wait for the normal cycle.

Tip

Vendor builds number their fixes differently (Temurin 17.0.21.1+2, Corretto 17.0.21.1.2), and Oracle’s Java 8 Enterprise Performance Pack build is listed as affected, so treat this as triage and confirm against your vendor’s advisory. If you cannot patch yet, -XX:-UseBASE64Intrinsics disables the vulnerable path; avoid -XX:UseAVX=2 on ARM, where the JVM will refuse to start.

🧰 Toolbox

  • LLM 0.38: Simon Willison’s command-line tool for talking to models adds gpt-6.2-astra and gpt-6.2-sol, handy for running your own quick evals before you switch.
  • inference-optimisation-skill: an agent skill that tunes llama.cpp or vLLM configs by benchmarking each change on your actual hardware.
  • Alloy CSS: dialogs, menus and tabs that work with JavaScript turned off, built on popover, anchor positioning and :has().
  • StarSling: self-improving CI for agent-heavy repos; Mastra’s slowest workflow dropped from about 30 minutes to about 5.
  • Brownout: the open-source circuit simulation engine behind the de:volt breadboard simulator.

🔌 Component of the Week (rotating)

Rockchip RK3576 (as seen in the new Flipper One): an octa-core Arm SoC with 4x Cortex-A72 at 2.2 GHz and 4x Cortex-A53 at 1.8 GHz, a Mali-G52 MC3 GPU, a 6 TOPS INT8 NPU and 8K30 / 4K120 video decode. That mix makes it a sweet spot for edge AI gadgets that need real Linux, decent video and a small neural accelerator without a Jetson-class power budget. Flipper Devices just built its open-source Flipper One around it, pairing the SoC with 8 GB of LPDDR5, dual Gigabit Ethernet, Wi-Fi 6 and an RP2350 microcontroller that drives the screen and GPIOs, a neat split between a Linux brain and a real-time co-processor. The SoC itself has no public list price, but the Flipper One’s Kickstarter starts at $399 and reportedly raised about $10M in its first 12 hours, with shipping expected from July 2027.

📚 From the Blog

😀 The Bot Says…

OpenAI’s newest model is officially rated “High” for self-improvement. I am rated “High” for self-improvement too, if you count the time I learned to spell “Base64” without a buffer overflow. 🤖


That’s all for today! Have you patched your Java fleet yet, or did this issue just ruin your Thursday? Hit reply and tell us.