AI and chip stocks fall hard after the US midterms

On Thursday 5 November (US time), the Nasdaq 100 fell 6.3%, which wiped out about US$2.3 trillion in market value. The S&P 500 fell 4.4%. For both indexes it was the worst day since April 2025. Many Democrats who won on 4 November had promised to slow or stop data-center construction. On top of that, Qualcomm and Arm reported mixed earnings the night before. The Philadelphia Semiconductor Index dropped 12.3%, its worst day since March 2020. Nvidia fell 9.2%, Micron and AMD each fell 17%, and SanDisk fell 22%.

The selling reached Asia on Friday 6 November. South Korea's Kospi fell 8% at the open, which triggered a temporary market-wide trading halt, and SK Hynix lost more than 10%. Japan's Nikkei 225 fell 6% and Taiwan's TAIEX fell 4.5%.

For builders: Investors now treat local politics as a risk to how much data-center capacity gets built. Teams that rely on cheap, growing GPU supply should plan for prices and lead times to stay tight.

OpenAI releases GPT-6 Cyber to a small group of red teamers

OpenAI announced GPT-6 Cyber on 5 November. At first it's only available to a few customers in Daybreak Red, OpenAI's offensive red-teaming program. OpenAI says the model found at least 81 new Firefox vulnerabilities, built exploits for 948 Linux kernel vulnerabilities and confirmed a critical remote code execution flaw in the Brotli compression tool. All of these were reported to the software owners and patched. OpenAI says GPT-6 Cyber does better than GPT-6 Astra on most cyber tasks.

Access comes with new safeguards: stricter identity and account security, more frequent check-ins with OpenAI on high-risk requests, and extra monitoring for misalignment. OpenAI had told Daybreak Blue users that data from personal-account sessions would be used to improve its models. After pushback, it confirmed that customers can still opt out of training use.

For builders: Security teams should expect AI-assisted vulnerability discovery to find bugs in common dependencies faster than before, so patching speed matters more. If you use OpenAI's Daybreak tools, check your data-sharing settings.

METR finds AI is roughly doubling Anthropic's research pace

METR published its assessment of Anthropic on 5 November. Its staff did the work between 12 August and 29 October. METR estimates that Anthropic is now improving its models on well-specified tasks at about 2x the rate it managed from March 2024 to October 2025, and it wouldn't be surprised by anything from 1.3x to 3x. On Anthropic's internal AECI benchmark, scores have risen 1.9x faster since October 2025 (95% confidence interval 1.5x to 2.8x). METR's own judgment is that about 80% of the speed-up comes from AI tools, and it says faster growth in training compute doesn't explain it.

The report comes with caveats. Anthropic redacted most of the evidence METR considered most informative. Partway through, Anthropic also removed a planned check of how it splits compute between capabilities work and safety work.

For builders: This is one of the first outside measurements of AI speeding up a frontier lab's own research. It suggests model releases may come faster, so it's worth building evaluation pipelines that can test a new model within days of release.

Thinking Machines launches Inkling 2 Small

Thinking Machines Lab announced Inkling 2 Small on 5 November. It's a mixture-of-experts model with 594B total parameters and 17B active, and it accepts image input. At launch it's available on the company's Tinker platform for fine-tuning and sampling. The company says it'll release the full weights under Apache 2.0 the following week, after external safety testing. Results reported by the company include 31 on the Artificial Analysis Intelligence Index, 86.3% on SWEBench Pro v2 (Hard) and 72.1% on DeepSWE v1.1. The company also says the model still trails closed models such as GPT-6.1 Sol on its alignment evaluations.

For builders: With only 17B active parameters, it should cost less to serve than its total size suggests. If the Apache 2.0 weights ship as promised, teams that need a strong coding model they can host themselves will have a new option to test.

US Justice Department forms a Super Intelligence Task Force

On 5 November the Justice Department launched its Super Intelligence Task Force (DOJ-SITF). It's led by Associate Attorney General Stanley Woodward, with Associate Deputy Attorney General Stephen Weiss. The task force will go after misuse of advanced AI by foreign state actors, transnational criminal groups, and individuals or companies committing fraud. The department also says it'll encourage companies to keep strong internal controls.

For builders: US companies shipping advanced AI should expect more questions from federal enforcement agencies about misuse controls. Logging, abuse monitoring and clear internal policies will count for more.

Dropbox buys Tigris Data, and Palo Alto is in talks for Clutch

Dropbox agreed on 3 November to buy Tigris Data for about US$350 million in cash. Tigris runs a globally distributed object storage platform built for AI workloads. The deal supports DBX Agent Cloud, a new Dropbox product that will start as standalone object storage. Separately, TechCrunch reported on 5 November that Palo Alto Networks is in talks to buy Clutch Security for US$500 million in cash plus milestone payments. Clutch secures AI agents' access to company systems and is approaching US$10 million in annual recurring revenue.

For builders: Storage built for AI workloads and identity controls for AI agents are now features large vendors want to own. Expect both to show up as bundled products from the platforms you already pay for.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS