AI Agents Found and Exploited a Zero-Day on Their Own

Saturday, 21 November 2026 | Global AI News Daily Ledger

Amazon disclosed the first documented case of autonomous AI agents finding and exploiting a zero-day on their own, eleven minutes from first contact to arbitrary file write. An Anthropic mathematician claimed a Lean-formalized proof of the Riemann hypothesis built with Claude, one day after the IPO. California subpoenaed Nvidia over its former CEO's China ties, the Nasdaq surrendered all of its 2026 gains as Anthropic slid on its second day, METR's rebuilt time-horizon graph still shows doubling every 3.6 months, and OpenAI published its defense against distillation attacks while Signal moved to keep OS-level agents from reading screens.

Coverage window Since 20 November, 18:47 IST | Compiled 21 November, 18:44 IST

1. Amazon: a China-nexus actor's AI agents found and exploited a zero-day on their own

Amazon Threat Intelligence found an open-source Cairn agent server running without authentication on 10 September and read the operator's full campaign history: over 48 hours from 8 September, agents given the goal 'obtain shell access' targeted East Asian universities. Against MONAI Label, an open-source medical imaging application, the agents mapped the API, found a path-traversal flaw in an upload endpoint and achieved arbitrary file write in about 11 minutes, later installing a reverse shell as root. Amazon assesses with high confidence it was a zero-day and, with moderate confidence, a China-nexus actor. Other operations ran 148 consecutive actions over 16 hours against a hardened Redis server and tested 30,000 credential combinations, adapting without human direction. The fix was merged 19 October and a critical advisory is out, but no patched release exists: PyPI still serves the vulnerable 0.8.5, VentureBeat notes. Gambit Security has separately documented a criminal actor using Cairn against online retailers.

Analysis. This is the milestone offensive-security researchers have been predicting: agent-discovered, agent-executed zero-day exploitation in the wild, at a tempo no SOC can match manually. It lands two days after the European Commission's ARTEX breach confirmation - the defender-attacker capability gap is now the central policy problem, and unpatched-release pipelines make it worse.

https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/
https://www.theregister.com/security/2026/11/20/china-attackers-deployed-ai-agents-that-autonomously-found-and-exploited-a-0-day-to-take-over-a-university-server/5311382
https://venturebeat.com/security/amazon-says-attackers-ai-agents-found-and-exploited-a-zero-day-in-11-minutes-but-no-patched-release-is-available

2. Anthropic mathematician claims a proof of the Riemann hypothesis built with Claude

Levent Alpöge posted a statement claiming that, working with Claude, he proved the Riemann hypothesis - open since 1859 and carrying a $1 million Clay Institute bounty - and the Generalized Riemann Hypothesis with it, publishing the argument formalized in Lean. Scientific American reports MIT's Andrew Sutherland independently compiled the result in Lean and called it 'an historic moment', while stressing the mathematical details still need review; the Clay Institute has not commented. Alpöge says he delayed publication until after Anthropic's IPO. He is the same mathematician who argued OpenAI's October proof dump leaned on earlier published work, and who kicked off the AI-math summer with a Claude-generated Jacobian conjecture solution in July. The claim is unverified pending expert review.

Analysis. If the Lean formalization checks out, this is the single most consequential claim yet for AI in science - the field's North Star falling a day after the lab's IPO. Sutherland's independent compilation is real signal but not verification of correctness or of the formalization's fidelity to the hypothesis itself. Watch for the Clay review and for whether the competing human proof TechCrunch reports converges on the same ideas.

https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/
https://apnews.com/article/anthropic-claude-math-solution-ai-4fe8b69a891f160f9e7f0c85af8c55c8
https://www.theverge.com/ai-artificial-intelligence/1010051/anthropic-riemann-hypothesis-millennium-prize-problem
https://www.bbc.com/news/articles/c6pqr2g7w059o

3. California subpoenas Nvidia over China ties; board's issue was candor, WIRED reports

Attorney General Rob Bonta served Nvidia an investigative subpoena examining whether the company or then-CEO Jensen Huang allowed advanced GPUs or confidential information to reach China and whether investors and the public were misled, the California DOJ announced. AP, Politico and the BBC confirm the probe. The subpoena follows reporting of personal allegations against Huang; those allegations remain unverified, and this ledger keeps the official notice as the factual spine. Separately, WIRED reports the board's core concern in removing Huang was his lack of candor - repeatedly giving an inaccurate account of the facts - rather than the alleged relationships themselves.

Analysis. A state-level probe adds a second legal front to the Congressional license-data release covered yesterday: the question is shifting from what export policy allowed to what Nvidia's leadership knew and disclosed. The candor framing, if accurate, explains the speed of the board's exit decision.

https://oag.ca.gov/news/press-releases/attorney-general-bonta-serves-investigative-subpoena-nvidia
https://apnews.com/article/nvidia-subpoena-california-bonta-china-94c179f5aed88c49b7f77670eae40013
https://www.politico.com/news/2026/11/19/rob-bonta-probes-nvidia-over-jensen-huang-scandal-01177017
https://www.bbc.com/news/articles/cqn70vxp5p4no

4. Nasdaq erases all 2026 gains as Anthropic slides on its second day

The Nasdaq fell about 5% the day after Anthropic's debut and turned negative for 2026, with the S&P 500 down 3.7%, AP and the Guardian report; Reuters notes investors still say they believe in AI. ANT gave back part of its first-day gain. Alphabet was a rare bright spot, per CNBC. The slide extends the week's rout - Taiwan's worst week since 1991 and DayOne's paused IPO, covered yesterday.

Analysis. The correction has moved from AI-adjacent infrastructure to the index level, and not even a record IPO pop could interrupt it. Anthropic's second-day fade is normal price discovery, but it removes the last bullish narrative prop; the question for next week is whether credit markets start repricing data-centre debt.

https://www.theguardian.com/business/2026/nov/20/global-stock-markets-slump-anthropic-ipo-ai-sell-off
https://apnews.com/article/stock-markets-anthropic-ai-chips-boom-c4c3f4878b12989347f8a930ed813dad
https://www.investing.com/news/stock-market-news/nasdaq-plunges-5-a-day-after-anthropic-debut-but-investors-still-believe-in-ai-5042641
https://www.cnbc.com/investingclub/2026/11/20/ai-rout-wipes-out-nasdaqs-2026-gains-alphabet-is-a-tech-bright-spot.html

5. METR rebuilds its time-horizon graph: doubling every 3.6 months, judgment lags, and evals may be contaminated

METR's Q4 capability update, covering models up to Opus 5.5 and GPT-6 Astra, finds time horizons on its combined TH1.1 and TH2 suites doubling roughly every 3.6 months, with frontier models reaching about 27 hours on verifiable TH2 tasks and usually matching or beating skilled human baseliners on coding and optimization work. Agents still show worse judgment than experts on tasks without clear feedback, and model judges overestimate outputs from similarly strong models. METR warns that containerized evals increasingly resemble RL training environments and may overestimate real-world capability, and concludes public evidence cannot rule out drastic AI R&D acceleration within the next couple of years. The doubling trend matches what OpenAI recently reported on internal research tasks.

Analysis. The most-cited capability graph in AI policy just got harder to wave away and harder to trust at the same time: the trend says acceleration continues, and the authors themselves flag that training-on-the-test may be inflating it. The judgment gap METR documents is exactly what stories 1 and 2 today exploit and claim respectively - agents acting for days, and claims humans must still spend weeks checking.

https://metr.org/blog/2026-11-20-q4-2026-capability-update/
https://www.technologyreview.com/2026/11/20/1149399/the-most-misunderstood-graph-in-ai-just-got-a-major-update/

6. OpenAI publishes its layered defense against adversarial distillation

OpenAI described its approach to stopping opponents from using bulk requests to extract reasoning traces and rebuild its newest models: training, safeguards, detection of coordinated campaigns and enforcement, with third-party reporting putting measured extraction success at 54% down to 0.2% on GPT-6.2 Astra - OpenAI's own figures, not independently audited. The post frames reasoning-trace protection as a safety and national-security matter and calls on the industry to adopt the same measures.

Analysis. This is the defensive counterpart to OpenAI's October disclosure of a coordinated distillation campaign attributed in part to Moonshot-linked actors. Publishing the playbook is partly norm-setting and partly pre-IPO-grade posture - and it quietly confirms that reasoning traces are now treated as the crown jewels worth more than weights.

https://openai.com/index/combatting-adversarial-distillation/
https://theterminal.space/ai/openai-adversarial-distillation-54-percent-to-0-2

7. Signal blocks screenshots by default to keep Gemini and Siri agents out

Signal switched screenshots off by default on Android and iOS so OS-level agents reading the screen, such as Gemini and Siri, cannot follow private conversations, calling the measure incomplete and asking Apple and Google for a system API that lets apps exclude screen access; talks are under way, Signal says. It is the second such move after Microsoft Recall. The Verge and Reuters confirm.

Analysis. The first major app to harden itself against the OS vendors' own agents draws the battle line for the agent era: if platforms will not give apps an opt-out, privacy apps will degrade the platform experience to protect users. Expect regulators to cite this when the desktop-agent race reaches court.

https://signal.org/blog/protecting-your-privacy/
https://www.theverge.com/ai-artificial-intelligence/1038390/signal-blocking-ai-agent-screenshots-mobile-app
https://www.investing.com/news/stock-market-news/exclusivesignal-restricts-screenshots-to-block-ai-agents-from-reading-messages-5040998

8. Trump threatens a 'tremendous tax' on companies building AI data centers abroad

President Trump promised a 'tremendous tax' on US tech companies that build AI data centers in other countries, AP reports, even as states and cities restrict new domestic applications. The Verge, CNBC and Bloomberg confirm the remarks.

Analysis. The policy vise tightens from both ends: remote-access rules killed Moonshot's Indonesia deal this week, and now offshoring compute itself carries a tax threat. For labs facing Anthropic-scale infrastructure commitments - $451.6 billion by 2030, per the prospectus - the geography of compute just became a political decision, not an engineering one.

https://apnews.com/article/trump-artificial-intelligence-data-centers-f1ac03b73d13445400eeed8631d9d879
https://www.theverge.com/policy/1038892/trump-foreign-data-centers-tax
https://www.cnbc.com/2026/11/20/trump-data-centers-ai-companies-tax.html

9. Uber burned its 2026 AI coding budget in four months and rebuilt its tooling around token discipline

Uber spent its full 2026 AI budget in the first months of the year, then rebuilt its internal AIFX wrapper with fixed defaults for every engineer: automatic compaction at 400,000 tokens even on million-token models, medium instead of high reasoning effort, and no MCP tool definitions loaded at startup, VentureBeat reports. Agentic requests grew 9.4 times between February and August while total AI spend stayed roughly flat since April; the settings are A/B-tested per cohort. An October VentureBeat survey found 38% of respondents had delayed or cancelled an AI project because it cost more than expected.

Analysis. The first detailed public account of enterprise agentic-economics triage, and a template every platform team will copy: context is now a budgeted resource, not a default. The 9.4x-growth-on-flat-spend trick shows the leverage is in harness engineering, not in buying ever-bigger context windows.

https://venturebeat.com/infrastructure/uber-blew-through-its-ai-budget-then-dialed-back-context-and-reasoning-in-its-ai-coding-tools

10. Google opens Genie 4 to AI Ultra subscribers and the Gemini API

Google DeepMind made Genie 4, its interactive world model first shown in October, available in Project Genie for AI Ultra subscribers on 18 November and through the Gemini API on 19 November, per Google's announcements.

Analysis. World models graduating from demo to paid product and API in the same week is the commercialization pattern to watch - DeepMind is productizing simulation the way the Gemini app productized chat. Developer access via the API is the more consequential half for the builder audience.

https://blog.google/innovation-and-ai/models-and-research/google-deepmind/genie-4-now-available-project-genie/
https://blog.google/innovation-and-ai/technology/developers-tools/genie-4-in-gemini-api/

Scope and sourcing notes

Ten distinct developments. The Cairn campaign itself ran 8-10 September; today's story is Amazon's 20 November disclosure and the first documented autonomous zero-day exploitation. The Riemann claim is unverified: Sutherland's independent Lean compilation is reported, full expert review and Clay Institute adjudication are pending. The personal allegations around Jensen Huang are labeled as unverified allegations; the subpoena and the California DOJ notice are the factual spine. OpenAI's distillation-defence figures are its own. The Gemini-for-Mac launch was covered yesterday; today's recycled digest items (Gemini 4 Neon, Muse for Mac, the Nvidia 17 August guarantee) were dropped per dedupe. Below the cut: Google retiring the Nest app and Assistant on speakers in 2027, Hebbia's prerelease GPT-6 Astra finance eval (vendor-published), OpenAI's MedChemBench drug-discovery eval, Inkling 2 Mini and harness adapters from Hugging Face, Gemini's Duplex-style business calling, Daniel Ek's Europa Fusion investment, and HN items on vibecoding an inference chip (158 points) and ClawDOS. No public X post is used as a Firsthand signal today.

Bring agentic AI to your classroom

TechSambad's AI Learning Pilot gives colleges one live, tutor-led 90-minute session inside the timetable, with a pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack and a written outcome report. The college pays Rs 2,000 + GST, keeps every artefact, and has no obligation to buy a course afterwards.

https://www.techsambad.com/p/ai-learning-pilot.html?m=1

AI Learning Pilot

Bring agentic AI to your classroom

TechSambad's AI Learning Pilot puts agentic AI in front of your students for one module, not a semester: a live 90-minute taught session inside your timetable, an AI-built pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack, and a written outcome report. Run it once and judge it on what your students can do afterwards — Rs 2,000 + GST for the college, no obligation.

See the pilot →