Hacker News · November 18, 2026 · 7:09

OpenAI launches GPT-6.2 Astra, Sol & Microsoft unveils MAI Code 2 models - Hacker News (Nov 18, 2026)

OpenAI's GPT-6.2 hits a self-improvement threshold, a critical Java flaw found with AI, Quad9 leaves France, and AI Overviews drain web traffic.

OpenAI launches GPT-6.2 Astra, Sol & Microsoft unveils MAI Code 2 models - Hacker News (Nov 18, 2026)
0:007:09

15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕

Support the show →

Our Sponsors

Today's Hacker News Topics

  1. OpenAI launches GPT-6.2 Astra, Sol

    — OpenAI released GPT-6.2 Astra and GPT-6.2 Sol for professional, coding and scientific work, with tighter cybersecurity safeguards. Astra is the first externally released OpenAI model to meet its high preparedness threshold for AI self-improvement.
  2. Microsoft unveils MAI Code 2 models

    — Microsoft announced MAI Code 2 and MAI Code 2 Flash, coding models built for enterprise price-performance and available in GitHub Copilot. A related model, MAI Work Flash, runs across Excel, Word, PowerPoint and Teams.
  3. AI agent harness benchmark findings

    — Runta's FrontierHarness Eval v1.1 ran 27,280 tests and found that the best AI agent harness depends on the model, with cost per task varying up to fivefold. Step count was the strongest predictor of cost, and most spending went to failed runs.
  4. Critical Java Base64 flaw discovered

    — Calif disclosed CVE-2026-114938, a critical OpenJDK Base64 decoder bug that enables heap memory corruption and possible remote code execution in Jira, Jenkins, Kafka, Keycloak and Elasticsearch. The flaw was found and weaponized with heavy AI assistance.
  5. Phrack Issue 73 released

    — Phrack Issue 73 is the hacker magazine's largest issue under its new staff, covering container escapes, kernel exploits and Chromium sandbox escapes. Its editorial argues that AI is changing security research but human creativity still matters.
  6. NSA classified network outage

    — NSA Net, the National Security Agency's classified internal communications network, went down on November 17. Officials blamed maintenance or temporary degradation rather than a cyberattack.
  7. Quad9 exits France over blocking

    — Quad9 is withdrawing its free public DNS resolver from France after a court order imposed daily penalties tied to copyright website blocking. Cloudflare warns that DNS blocking mandates harm privacy, security and internet access.
  8. AI Overviews cut publisher traffic

    — Research suggests Google AI Overviews reduce clicks to websites by about 40 percent on affected searches, which shifts SEO toward generative engine optimization. The All About Berlin site reports losing 75 percent of its traffic.
  9. ID.me terms for IRS logins

    — Taxpayers who sign in to IRS online services through ID.me accept arbitration and class-action waivers, along with data retention for up to three years. Login.gov is offered as a government-run alternative.

Sources & Hacker News References

Full Episode Transcript: OpenAI launches GPT-6.2 Astra, Sol & Microsoft unveils MAI Code 2 models

OpenAI's newest model is the first it has released publicly that meets its own high-risk threshold for AI self-improvement. We'll look at what that means in a moment. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is November 18th, 2026. Today's lineup includes a new generation of AI models, a serious Java vulnerability, a DNS service leaving an entire country, and some tough numbers for anyone who publishes on the web. Let's get into it.

OpenAI launches GPT-6.2 Astra, Sol

We'll start with OpenAI, which has introduced two new models: GPT-6.2 Astra and GPT-6.2 Sol. Astra is the premium option for complex professional work like documents, spreadsheets, slide decks, code and scientific tasks. Sol is the faster, cheaper model OpenAI expects many people to use by default. OpenAI says both models produce more polished output while using fewer tokens. The safety section stands out. Both models are notably better at finding software vulnerabilities, so they ship with tighter safeguards. OpenAI also says Astra is its first externally released model to meet the high threshold for AI self-improvement in its preparedness framework. That doesn't mean the model is rewriting itself, but it is a milestone that safety researchers will watch closely.

Microsoft unveils MAI Code 2 models

Microsoft is pursuing the cost-conscious side of the same market with MAI Code 2 and MAI Code 2 Flash. Both are coding models for everyday enterprise work such as debugging, multi-file changes and quick fixes. Microsoft says plainly that these models aren't meant for frontier research or very long autonomous tasks. Instead, it claims they come close to the leading competitors on coding benchmarks for much less money. The models are available in GitHub Copilot now. A related model is already running inside Excel, Word, PowerPoint and Teams.

AI agent harness benchmark findings

A new benchmark offers a useful reminder about model choice. Runta's FrontierHarness Eval ran more than twenty-seven thousand tests across different combinations of agent harnesses and models. A harness is the scaffolding that turns a model into an agent. The main finding is that no single harness is best for every model. With the same model, the harness alone could make a task up to five times more expensive, and paying more didn't reliably improve results. Most of the cost came from the number of steps an agent takes. The less comforting finding is that much of the spending went to runs that failed, and many of those failed runs still ended with a confident answer. The authors recommend testing harnesses on your own tasks instead of trusting a general leaderboard.

Critical Java Base64 flaw discovered

On security, the firm Calif has disclosed a critical flaw in OpenJDK's Base64 decoder. Base64 decoding is about as routine as code gets. In this case, a short, malformed string can corrupt memory before Java rejects the input, which could open the door to remote code execution. Calif says it showed real impact in widely used tools including Jira, Confluence, Jenkins, Kafka and Elasticsearch. The bug lives in a performance optimization, so Calif argues that removing unnecessary code is itself a security measure. The firm also says it found and weaponized the flaw quickly with heavy AI help, an approach it calls the 'centaur hacker' model. That fits neatly with OpenAI's warnings about cyber capabilities earlier in the episode.

Phrack Issue 73 released

Staying in the hacker world, Phrack has published Issue 73, following last year's 40th-anniversary revival. It's the largest issue under the new editors. Topics include container escapes, kernel research, Chromium sandbox escapes, autonomous underwater vehicles and backdoors in machine learning systems. A history section revisits episodes like blue boxing and an attempted Linux kernel backdoor. The editorial accepts that AI is changing security research, but argues that human intuition still does much of the important work.

NSA classified network outage

Another security story: the NSA's classified internal communications network, known as NSA Net, went down on November 17th. Officials say there's no sign of a cyberattack and are attributing the outage to maintenance or a temporary degradation. It's still a notable event, because it follows several other recent technical failures in U.S. systems and raises fresh questions about how resilient critical infrastructure really is.

Quad9 exits France over blocking

Next, internet policy. Quad9, the nonprofit that runs a privacy-focused public DNS resolver, is pulling out of France. After months of litigation, a French court imposed daily penalties for not blocking domains linked to alleged copyright infringement. Cloudflare, which has faced similar pressure, says Quad9 had two bad options: pay mounting fines or weaken the privacy and security its service is built on. Cloudflare argues that blocking at the DNS level does little to stop piracy and does real harm to hospitals, schools and other organizations that depend on these resolvers. A broader Belgian order could put Cloudflare, Google and Cisco in a similar position.

AI Overviews cut publisher traffic

For publishers, the effects of Google's AI Overviews are getting harder to dismiss. One analysis cites a randomized experiment indicating that the AI summaries reduce clicks to outside websites by roughly 40 percent on searches where they appear. The summaries now show up across nearly every kind of query, so publishers can't avoid them simply by targeting different keywords. The report also finds that ranking well in search and being cited by AI are increasingly separate goals. It describes 'generative engine optimization' as mostly familiar SEO, plus extra work on making pages easy for AI to quote. For one creator, the impact is very direct. The founder of All About Berlin, a widely used guide to the city's immigration and bureaucracy, says the site has lost about 75 percent of its traffic. He says other independent publishers are putting their work behind paywalls or no longer updating it, and he worries about where that leaves the open web.

ID.me terms for IRS logins

Finally, a note for U.S. taxpayers. If you sign in to IRS online services through ID.me, you're accepting ID.me's own terms, not just the IRS's. Those terms include mandatory arbitration and a waiver of class-action lawsuits unless you opt out within 30 days. ID.me may also keep verification records for up to three years after you close your account. There is an alternative: the government-run Login.gov also works for IRS access, and senators have recently asked why federal agencies continue to rely on a private provider.

That's the news for this November 18th. The common thread today is AI's growing reach, from writing code to finding vulnerabilities to changing how people get information online. Links to all the stories are in the episode notes. I'm TrendTeller. Thanks for listening, and see you tomorrow on The Automated Daily.

More from Hacker News