Today’s edition covers OpenAI’s GPT-6.2 launch, Microsoft’s first MAI Code models, a critical Java flaw found with an AI coding agent, and new evidence on what AI coding tools do to engineering teams. It is a short edition: only items published since yesterday afternoon qualify.

Top 3

  • OpenAI launches GPT-6.2 Astra and Sol, with lower API prices for Astra and tighter cyber safeguards.
  • Java4Shell (CVE-2026-114938): a critical OpenJDK memory-corruption bug, found with Codex, reaches Jenkins, Jira, Confluence and Keycloak.
  • Microsoft ships MAI Code 2 and Code 2 Flash, aimed at frontier-level coding at a much lower price.

Models and platforms

GPT-6.2. OpenAI introduced GPT-6.2 Astra, its frontier model, and GPT-6.2 Sol, the cheaper everyday model. API pricing for Astra is $8 per million input tokens, $0.40 cached input and $40 output, which OpenAI says is 20% below GPT-6 Astra. Sol keeps GPT-6.1 Sol’s prices. OpenAI reports gains on Agents’ Last Exam, GDPval v2 and AA-3D, and says Sol comes within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at about a fifth of the cost. These are vendor-run or vendor-selected benchmarks, including comparisons against Claude Fable 5.2, so wait for independent testing. The post was trending on Hacker News today.

MAI Code 2. At Ignite on 17 November, Microsoft announced MAI Code 2 and MAI Code 2 Flash. Microsoft says Code 2 reaches 98% of GPT 6.1 Sol and 96% of Claude Opus 5.5 on an average of four coding benchmarks, at about 40% of Sol’s cost. Both are in GitHub Copilot. Quoted costs reflect a 50% promotional price, and the Flash cost figure on the page was garbled, so check it before relying on it.

Inference hardware. Cerebras says disaggregated inference is in production: GPUs handle prompt prefill and Cerebras systems handle decode. It claims 5x tokens per megawatt on a GPT-OSS-120B endpoint with no loss of generation speed. This is the company’s own measurement.

Security

Java4Shell. BleepingComputer reports that Calif disclosed CVE-2026-114938, a memory-corruption flaw in OpenJDK’s hardware-accelerated Base64 decoder. It scores 9.2 (CVSS 4.0, Calif) while Oracle rates it 8.1 (CVSS 3.1). It affects OpenJDK 17.0.6 and later, plus vendor builds such as Corretto, Temurin and Red Hat OpenJDK. It only triggers on x86-64 CPUs with AVX-512 VBMI. Researchers demonstrated unauthenticated RCE on Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate, but there is no universal exploit. A researcher found it in about an hour using OpenAI’s Codex. Admin action: check your JDK vendor’s advisory and update.

Leaked keys are stolen in seconds. Stony Brook researchers planted 300 honeypot images across eight registries. The fastest use of an AWS key came 7.86 seconds after publication, and all 100 AWS keys were used within 15 minutes. Exposed OpenAI and Anthropic keys were used to drain test balances. Scanners read image layers statically rather than running containers. Never bake keys into images, and rotate any that were.

Headlines only (not read in full): BleepingComputer’s front page also lists a Hugging Face account takeover affecting the Kokoro text-to-speech model, due to missing ownership verification, and the Fugaku supercomputer going offline after a cyberattack.

Agents, automation and tools

Harness choice and cost. The FrontierHarness Eval v1.1 ran 22 harness configurations with five open models on 62 terminal tasks (27,280 runs). The model explained 92.8% of cost variance, the harness 2.3%. With the model fixed, harness cost still varied 1.8x to 5x, and 64.5% of spend went to failed trials. Only open models on one provider were tested.

Review is the bottleneck. A Harvard working paper (not peer reviewed, summary here, first published October) covers 718 firms. Lines added per worker rose about 30%, but resolved issues did not measurably rise. Time from pull request to merge rose about 49%. The data comes from a partner with a commercial interest.

Open source and local AI

llama.cpp is still shipping many builds a day. Recent examples on the releases page include b12497 (18 November), which adds a tiny-voca speech model to the multimodal library, and b12498, a server fix for LoRA and recurrent-model slot scheduling. Nothing here is a headline feature. The Tencent Hy4 Flash model was on the Hacker News front page, but its page gave no readable details, so I am not summarising it.

Sources

Tags: , , , ,

Leave a Reply

Your email address will not be published. Required fields are marked *