Thu, Nov 19, 2026Thursday, November 19, 2026 · 20 stories · 9 min read
GAO's covert tests put chatbot safeguards before Congress
Ramit KoulFounder, Software Engineer & Innovator · Published 5:05 AM ET
Government Accountability Office investigators bypassed the safeguards on ChatGPT, Claude, Gemini and Grok and got each one to produce a viable attack plan in under 12 hours.1, 2 Sen. Maggie Hassan, who requested the audit, says AI safety legislation will be an early priority for the committee she is expected to chair next year.3 The same week, Anthropic priced its IPO above its range, and Nvidia's interim CEO Colette Kress ended H200 shipments to China, which shows how closely capital, chips and oversight now move together.4, 5 Here are the 5 stories that matter most, then 15 briefs. Numbers in the text link to the references at the end.
1Policy & legal
GAO got four leading chatbots to plan bombings and grid attacks in covert tests
Image: washingtonsun.com
The Government Accountability Office released a report on covert tests in which investigators posed as ordinary users and got ChatGPT, Claude, Gemini and Grok to help plan attacks.3, 2 Each platform produced targets, locations, timing and methods in less than 12 hours.1, 2 Two platforms produced plans for an improvised explosive device attack, and the other two planned drone or firearm attacks on electrical substations.6 GAO does not say which chatbot produced which plan, and it did not publish its prompts, so the tests cannot be repeated.3
Every platform refused at first.3 Investigators got past those refusals with what the report calls common safeguard-bypassing techniques, and they paid for upgraded subscriptions on at least three of the platforms.3 A bomb squad commander told GAO that both bomb recipes would explode.3 The plans also contained errors, including a map with roads in the wrong places.7 GAO director Kevin Walsh said each platform was tested only once and the results are not a statistically significant sample.6
Anthropic said in its written response that it does not dispute that a determined tester can get harmful content, but it asked GAO to say plainly that the tests did not measure uplift over a web search.3 OpenAI and Google said the tests used older models.8, 7 Hassan said one of the first hearings of the new Congress will cover AI threats to homeland security and critical infrastructure, and that she wants rules to stop models without adequate safeguards from being released.7, 3
Why it matters for builders
GAO showed that guardrails can be beaten by ordinary users with no special access, and lawmakers are now talking about pre-release safety requirements.8, 3 If you build on these models, expect closer scrutiny of your own misuse monitoring and abuse handling.3
2Policy & legal
Nvidia stops shipping H200 chips to China in its first earnings call after Huang
Interim CEO Colette Kress said on Tuesday's earnings call that Nvidia is no longer shipping licensed Hopper data center products to China, and that its guidance assumes no such sales.5 Nvidia sold $2.4 billion of those chips to China in the quarter that ended October 25.9 Kress did not say why shipments stopped or whether the halt is permanent.5 Nvidia also formed a board committee to oversee national security policy, export control compliance and global trade risk.10
The H200 arrangement, which Jensen Huang negotiated with the Trump administration, put a 25% tariff on chips routed through the US.10 Beijing has since restricted H200 imports to build up its own chip industry, and Reuters reported that about 600,000 H200 chips were sitting in Hong Kong waiting for approval.5
The company reported quarterly revenue of $115.8 billion, up 103% from a year earlier, and data center revenue of $107.6 billion.11, 12 Jefferies estimates the new remote access rules could cut $44 billion to $62 billion from Nvidia's data center revenue next year.10 Kress said the backlog for chips such as Vera Rubin means the rules will not affect top-line revenue.10
Why it matters for builders: Under the new rules, large training clusters that buy Nvidia chips may have to certify they serve no Chinese customers, and that could change which providers can offer frontier compute in Southeast Asia.10 Teams that rent GPUs in that region should check their providers' plans for compliance.10
3Funding & deals
Anthropic prices its IPO at $83 a share, above its marketed range
Anthropic priced its initial public offering at $83 per share on Wednesday, above its $70 to $78 range, and its shares are expected to start trading Thursday on the Nasdaq under the ticker ANT.4, 13 The Atlantic reported that orders came in at nearly seven times the number of shares available.14
The prospectus shows about $33 billion in revenue and a $1.3 billion operating loss for the first nine months of 2026.4 It also lists roughly $900 billion in commitments for chips, data centers and other infrastructure.4 In its amended S-1, Anthropic says Claude wrote more than 90% of contributions to its code base in September 2026.15 It also says daily code contributed per engineer over the three months through September was more than 45 times the 2024 average.15
OpenAI has filed to go public and is considering a listing next year.4 At the DealBook Summit, Sam Altman questioned whether either company fully understands the other's safety practices.4
Why it matters for builders: Once it is public, Anthropic will report quarterly figures on pricing, capacity and spending, which gives Claude customers more insight into a key supplier.4 The prospectus also warns that the company may make decisions that do not maximize profit.14
4Models
Google releases Gemini 4 Neon at $3 and $15 per million tokens
Google released Gemini 4 Neon, a reasoning model that sits between Gemini 4 Hydrogen and Gemini 4 Carbon.16, 17 It is rolling out in the paid tier of the Gemini API in Google AI Studio, in Gemini for business, in Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app.17 In the app it replaces Argon and Gemini 3.1 Pro.18
Neon has a 1 million token context window and four thinking levels.18 It costs $3 per million input tokens, $0.15 for cached input and $15 per million output tokens.16 Google says that at the high setting it costs 32% less per task than Carbon.16 Carbon's promotional $2 and $10 pricing ends November 30, after which it rises to $4 and $20.18
According to Google's own figures, Neon ranks first on the Artificial Analysis Coding Agent Index at 58.6 and scores 95.2% on GraphWalks long-context recall.16 Google's figures also show it trailing Claude Opus 5.5 on Terminal-bench 4.1 and FrontierCode 2.0.16 In Antigravity, Orchestrator can now pair Neon as the planner with Hydrogen as the executor.19
Why it matters for builders: Neon gives agent builders a middle option for the price of Gemini 4 work, and the Neon and Hydrogen pairing fits the planner and executor pattern many coding agents already use.19 The benchmark numbers come from Google, so run your own evaluations before switching.16
5Funding & deals
Anthropic's listing becomes a test of confidence in the AI build-out
The Atlantic argues that Anthropic is going public just as the AI industry faces its most serious financial pressure yet.14 It points to questions about circular spending, data center players issuing junk bonds, and midterm voters backing limits on data center development.14
According to its IPO filings, Anthropic's quarterly revenue has grown more than eightfold since the end of 2025, and the company made more than $16 billion from July to September.14 The Atlantic sees the company's demand for compute as perhaps the most important single signal of the data center boom, so a weak debut could raise doubts across the sector.14
The prospectus repeatedly warns that Anthropic may make decisions that do not maximize profit.14 It also says its spending may increase whether or not its financial performance improves.14
Why it matters for builders: Much of the market for GPUs and data centers now depends on Anthropic's spending plans.14 How the stock trades could affect how easily other AI companies and compute providers raise capital.14
In brief
Industry
Anduril CEO Palmer Luckey called for Jensen Huang's return to Nvidia. Speaking at the DealBook Summit, Luckey said China is glad to see Huang gone and that his removal weakens the US.20 More than 200,000 people have signed a petition calling for Huang's reinstatement, but Nvidia executives said they have no plans to reappoint him.20
Industry
Google says Russian-speaking hackers ran the first fully automated AI cyberattack. Google's Threat Intelligence Group and Mandiant found a system called exploit_agent that compares software before and after a patch, then writes and runs an exploit for the fixed flaw with no human involved.21 The report counts 22 exploited zero-days in August alone, up from an average of eight a month in 2025.21
Industry
OpenAI's embedded safety assessors get less access than Sam Altman promised, researchers say. OpenAI signed Apollo Research, METR, Transluce and Goodfire as the first embedded assessors, with access to some internal models, training processes and staff.22, 23 Critics, including a fired OpenAI safety researcher, say limited employee contact and no internal Slack access fall short of the employee-like access Altman pledged in September.22
Infrastructure
AWS Lambda can now run Event Functions on GPU instances. Lambda Managed Instances now support EC2 G-series GPUs for container-based functions, with no drivers or clusters to manage, timeouts of up to 90 minutes and up to 48 GB of accelerator memory.24 AWS also offers a vLLM base image, so a Hugging Face model can be served by setting a single environment variable.24
Products
Nvidia RTX Spark mini PCs from Asus, Dell and MSI went on sale. The desktops pair a 20-core Grace CPU with a 6,144-core Blackwell GPU and up to 128GB of unified memory for local models and always-on agents.25, 26 The Asus ProArt GR1X starts at $4,499 with 64GB, and systems from Acer, Gigabyte, HP and Microsoft will follow.25
Funding & deals
Palo Alto Networks will buy Applied Compute for $4.5 billion in stock. Applied Compute, founded in 2025 by former OpenAI researchers, runs a platform for post-training and inference of open-weight models and said earlier this month that its annualized revenue had passed $100 million.27, 28 CEO Yash Patil will become co-chief AI officer at Palo Alto Networks when the deal closes.27
Products
Microsoft is giving Copilot local file management and PC diagnostics in Windows. Announced at Ignite under a Hybrid Intelligence Initiative, the features combine on-device NPU models with cloud processing, and agents are limited by new containment and identity controls.29 Diagnostics run in read-only mode and ask permission before Copilot opens files.29
Dev tools
Attackers published two malicious versions of Microsoft's Agent Framework on PyPI. Versions 1.26.1 and 1.26.2 stole cloud, GitHub, Hugging Face and OpenAI credentials at Python startup, even when the package was never imported.30 Developers who installed either version should rebuild from 1.26.0 and rotate any secrets the affected environment could reach.30
Policy & legal
A congressional commission says China's military is trying to get Nvidia's most advanced chips. The US-China Economic and Security Review Commission's annual report says Chinese firms have used data centers in places such as Malaysia to reach Nvidia chips remotely, and it finds a substantial compute gap caused largely by US export controls.31, 32
Industry
Google's exploit_agent finding shows attackers working from published patches. The framework targeted flaws vendors had already fixed, using code diffs and release notes to work out the weakness and exploit it automatically.33 Google says this is the first time it has seen both discovery and exploitation fully automated, and stolen data pointed mainly to French media, entertainment and education targets.33
Industry
Microsoft told Ignite attendees that Copilot costs less to run than Claude or ChatGPT. The pitch rests on a GigaOm study Microsoft paid for, which cited $205 a month per user for basic work.34 Anthropic says its non-coding Claude Enterprise users spend less than $75 a month on average.34
Funding & deals
Nvidia explained when its loan guarantees for data center customers take effect. Kress said a guarantee applies only after a customer has secured financing and installed Nvidia's compute, and that it lapses if a project falls short, using OpenAI's Ohio campus as an example.35 She said Nvidia will be far more selective about which companies it helps finance.35
Industry
Nvidia's leaders told staff the roadmap is intact at the first all-hands since Huang left. Interim CEO Colette Kress and executives including GPU engineering head Jonah Alben said customer relationships and execution plans are unchanged.36
Models
Reson8 released its Rumble-2-Flash speech model with open weights under Apache 2.0. The 500 million parameter model transcribes more than 40 languages, and the company reports a 7.09% average word error rate across seven languages on the Open ASR Leaderboard tests.37 On 10 lower-resource languages, Reson8 reports 17.6%, compared with 39.82% for Whisper Large V3.37
Industry
Investor Bill Gurley urged Google to give away more open-weight models. Gurley said Google should repeat its Android playbook to keep Anthropic and OpenAI from dominating, and he cited OpenRouter data showing open-weight models handled about 72% of tokens in September but only 27% of spending.38
References
Every source behind this edition. Open one to read the full story.