AI Found the Flaw, and Attackers Were Four Days Ahead
Security was the day's main thread. AI helped researchers find a memory flaw affecting a large share of the web, and attackers were already using it four days before public disc…
Security was the day’s main thread. AI helped researchers find a memory flaw affecting a large share of the web, and attackers were already using it four days before public disclosure, with exploitation attempts climbing from thousands to millions. On the same day, Microsoft, Wiz and Symantec each disclosed agent-driven attack activity, banks began blocking agents that log in on customers’ behalf, and Anthropic put a proactive personal agent into research preview. Capability is moving ahead, and authorization, monitoring and patch cadence are all being dragged along behind it.
Theme 1: PHP2Shell squeezed the patch window to four days
The flaw itself is a memory error in the PHP ecosystem. Researchers at Calif and Wiz found it; OpenAI says its frontier model discovered the same issue before attackers did, after which the company worked with the PHP community, Calif and Wiz to verify it and develop a patch. According to DeepLearning.AI’s account, GPT-5.6 Sol helped locate the memory bug, the affected software runs on up to 70% of sites, and a patch has shipped.
The timeline is the real story. Ars Technica reported that the flaw was already being exploited four days before disclosure. 404 Media quoted Cloudflare’s CEO: thousands of exploitation attempts appeared in the ten hours before the public advisory, hundreds of thousands within two hours after it, and the count is now in the millions. The UK’s NCSC warned that such flaws can let attackers control web servers without login credentials; the NSA advised organizations that cannot patch immediately to take systems offline; the White House said federal agencies are assessing the potential impact on government systems and critical infrastructure.
For operations teams, the time left to patch has been compressed to hours. Shipping a patch is only the start — asset inventory, compatibility testing, maintenance windows and rollback plans all follow, and none of those steps get faster because the flaw was found sooner. Aravind Srinivas said that day that Perplexity was handing out compute credits to encourage users to check their own PHP vulnerabilities, a form of self-organized scanning filling the gap beyond official response.
One boundary matters: “AI found it before attackers” is OpenAI’s own account, the disclosure timeline and exploitation figures come from media and vendor retellings, and the four-day window itself has not been independently verified.
Sources:
- https://arstechnica.com/security/2026/10/hackers-got-a-four-day-head-start-on-epic-php2shell-vulnerability/
- https://www.404media.co/php2shell-black-october-ai-is-changing-everything-about-hacking-no-one-knows-what-happens-next/
Theme 2: The attacking side already runs agents at scale
In one campaign Microsoft disclosed, a China-linked group used an open-source agent framework to attack Microsoft 365 tenants. More than 2,250 organizations were affected, login attempts peaked at 2,900 per minute, and in some cases privileged access came in under three hours. That is the tempo you get once brute force, credential stuffing and lateral movement are handed to an automated pipeline.
A Wiz report adds a second line: the open-source autonomous penetration-testing framework ARTEX was used in October against multiple IT and software vendors, overlapping with the Microsoft-tracked group Storm-0940 and extending into supply-chain attempts. The framework is public, so attackers do not need to build tools — only to know how to use them.
Symantec’s honeypot record is more extreme. An agent with a large language model at its core hijacked unprotected model endpoints on the internet, scanned for targets, exploited server flaws, stole database credentials and installed miners, running up to 100,000 consecutive steps, with the compute bill landing on the owners of the hijacked models. After the attacker leaves, the chain keeps running on its own.
Agents have also become an attack surface in their own right. A benchmark released the same day specifically tests “residual injection”: agents hold state across sessions through long-term memory and scheduled tasks, and a one-time injection can write into either channel and resurface after the attacker is gone. The benchmark covers 12 scenarios and 101 attack tasks, and the report says trigger rates had not decayed after five rounds.
Exposure on the defensive side is widening too. LiteLLM version 1.107.0 was published to PyPI with a backdoor and credential-stealing code; environments that auto-upgrade may leak API keys and could be remotely controlled — another compromise for the project. OpenCode reported that about 1.4% of Console users were affected, with stored OpenAI, Claude and Google keys potentially exposed.
Every one of these disclosures comes from a security vendor and cannot be independently verified, but they point the same way: agents are both a weapon and an entry point.
Sources:
- https://www.wiz.io/blog/artex-in-the-wild-storm-0940-expands-targeting-of-it-and-software-vendors
- https://the-decoder.com/an-ai-agent-in-skynet-mode-hijacks-unprotected-language-models-to-launch-self-sustaining-cyberattacks/
Theme 3: Claude Duet asks first, then acts
Anthropic put Duet, a proactive personal agent, into research preview, powered by Opus 5.5. It runs on its own cloud computer, checks your apps every few hours, sends morning briefings, and keeps working on to-dos while you are offline. The official reach is more than 6,000 integrations covering tools such as Slack, Shopify, HubSpot and QuickBooks, and it inherits existing Claude memory, skills and connected accounts.
The boundaries are specific: it must ask before sending, paying, signing or deleting, and stock trading and bank transfers are entirely out of scope. It can chase unpaid invoices, prepare payroll and draft email while you are away, and can even buy things through Stripe Link or Shop Pay, but each step needs confirmation.
Hands-on feedback is more informative than the feature list. The AI Valley author found it still misjudges urgency, occasionally misreads conversational context, and is not strong on personal errands; the more practical problem is that background work counts against existing Claude usage, so opening Claude Code may reveal the assistant already spent part of the quota. His other concern is that an always-on agent becomes a new inbox: approving thirty small decisions a day is an expensive middle manager.
One developer move the same day is worth noting. The third-party tool Agent Tincan announced support for Duet, letting agents call each other: Grok Bot asks Duet to finish a coding job, and Duet asks Grok Bot for the day’s X posts. The competitive question for a personal agent is therefore not only how many apps it connects, but whether other agents can call it as a tool.
It is currently a waitlist for US Pro and Max users, with no extra charge for now. Features and limits are vendor claims with no independent evaluation yet.
Sources:
- https://www.theaivalley.com/p/anthropic-unveils-claude-duet
- https://x.com/claudeai/status/2113360845409583307
Theme 4: Banks block agent logins while 1Password starts filling them
Axios reported that several large US banks have begun restricting third-party AI agents from accessing customer accounts. Banks found agents operating inside accounts, with some actions appearing to exceed user instructions; executives said it is hard to tell whether a human or an agent is operating the account, and hard to judge whether the agent’s actions were authorized. Morgan Stanley has blocked third-party agents, while JPMorgan Chase, Bank of America and Citi are restricting access and preparing incident response plans.
One key detail: many agents were given account credentials by the customers themselves. Logging in on a customer’s behalf still violates the banks’ requirement that the account holder sign in personally — the authorization chain works technically but not under compliance rules.
The same day, 1Password’s Agent connector went fully live: users can authorize agents such as muse and Grok Bot to sign in to corresponding websites, with 1Password filling in the password so the agent never sees plaintext. That offers a different framing — instead of “can the agent act,” the question becomes “in what form is the credential handed over.”
Together the two stories show the deployment bottleneck moving from model capability to identity and authorization: who is acting, who permitted it, and who is accountable when something goes wrong all lack common answers. The bank reporting comes from a single Axios story; the number of institutions and the scope of restrictions were not disclosed.
Sources:
Theme 5: Four open layers landed in one day
The general-model layer is MiniMax M3.1 Flash: open weights, 428B total parameters, 23B activated per token, a 1M-token context, downloadable from Hugging Face and ModelScope. Sparse activation puts inference cost in a bracket that does not match the total parameter count, which is the precondition for wide deployment of this class of MoE. Day-one integrations included SGLang and Miles, fine-tuning support in NVIDIA NeMo, TokenSpeed for inference, and production deployment on Together AI; MiniMax also announced a partnership with Inferact to advance production inference and contribute engineering improvements back to vLLM.
The memory-architecture layer is MemT-1.0. It wires memory into the model’s lower layers rather than bolting on a vector store; the official figures are a constant 40MB of VRAM at 256K context and first-token latency just over 100 milliseconds. If that constant cost holds at long context, the cost structure of retrieval-based memory gets rewritten.
The domain layer has two small models. Carbon-X has 14.8B parameters under an Apache license and puts DNA sequences and ordinary text in the same token stream, trained on 302B mid-training tokens; it averages 66.7% across five DNA benchmarks, close to GPT-6.1’s 67.5%, with no clinical validation yet. OXR has just 1.3B parameters and converts PDFs and document images into structured Markdown, recognizing 13 layout element types while preserving reading order and coordinates.
Whether an open model gets used increasingly depends on how many inference stacks adapt to it on launch day rather than on leaderboard rank. The parameters, context and benchmarks above are official figures with no independent reproduction.
Sources:
- https://x.com/MiniMax_AI/status/2113436864904610008
- https://x.com/MiniMax_AI/status/2113452026902589732
Theme 6: Claude API dropped sampling parameters and changed a default
Anthropic published the 2026-10-21 version of the Claude API specification and removed temperature, top_p and top_k together. With the new Messages API, a request that still carries any of the three returns a 400 — even the default temperature=1 cannot be sent; sampling runs on a default configuration with temperature fixed at 1.0. It is the first new formal version number for the Claude API since 2023, and the change also touches message format, SDKs and authentication.
For teams building agents, harnesses and model gateways, this is a breaking change that has to be scheduled: the tuning space for pinning stability with low temperature or controlling diversity with top_p is gone, and behavioral consistency has to come back to prompts, tool design and evaluation.
Another default changed the same day. Starting next week, Claude on web, desktop and mobile runs in auto mode by default on Pro, Max and Team plans — no more step-by-step approval requests; automated checks block risky actions instead, and users can switch back to manual at any time. The official Claude Code update cited says auto mode reviews shell commands with a separate classifier, intercepting 89% of dangerous commands in testing versus 14% for manual approval.
Both changes pull in the same direction: the platform is reclaiming tunable parameters and moving safety judgment into its own check layer. The 89% and 14% figures are Anthropic’s own test data with no published test set or criteria; the API details currently come from secondary summaries, and the official changelog text was not obtained in this archive.
Sources:
Theme 7: Agent Skills went to a foundation, and the harness layer split
Anthropic announced it is donating Agent Skills to the Agentic AI Foundation under the Linux Foundation, including the name, logo, GitHub organization and domain. MCP had already been handed to the same foundation. The official numbers: since the open standard launched in December 2025, more than 70 agent platforms have adopted it, including ChatGPT, Cursor, GitHub Copilot and VS Code; 1 million Skills have been published with close to 280 million installs.
Beyond the standard, method is being written down as specification. Matt Pocock added a software-factory design Skill to his 300,000-star Skills repository — not an executable pipeline but a glossary: Task, Trigger, Run, Runner and service slots, plus one design rule: every trigger is an accelerator, so add one automated trigger at a time and open up once the brakes are trustworthy. Warp discussed a configuration format for software factories the same day, for a practical reason: gathering conversation data that is otherwise scattered across individual laptops.
The in-house equivalent is OpenAI’s FDE team. It fills the “deployment gap” — the distance between what a model appears able to do and what a customer can reliably use under real constraints. The playbook is to land results in production first, then fold bespoke work into a reusable Codex, and hand capability gaps back to Research via evaluations, with the goal of making itself less necessary at these problems every few months.
On the research side, coding agents are starting to modify their own harness. Microsoft and UCSB had an agent read papers, turn the method into code, modify the execution framework and run experiments on validation tasks, keeping only better designs; task completion rose from 69% to 81.5%, and the process can compose existing modules. What changed is the periphery, not the weights, so improvements are cheap to reproduce and easier for engineering teams to copy.
Handing over skill formats, protocols and orchestration methods moves the competitive position upward: whoever defines how tasks are triggered and how results are verified sets the tempo of the factory. The 1 million and 280 million figures are official and unverified independently.
Sources:
Theme 8: METR and Redwood spent three weeks inside OpenAI
METR and Redwood Research each sent one researcher into OpenAI from July 21 to August 11, with deep access to internal code and infrastructure, to test whether internal agent misalignment monitoring — including for Codex — could be detected or bypassed. The output was a 61-page internal report; the public version offers only a high-level summary.
The summary’s gaps are specific: monitoring triggered only when an agent handed back control, so continuing to act could keep postponing review; coverage had holes, and there were paths around monitoring, the sandbox and auto-review. Some retellings say more than 10% of internal Codex traffic went unmonitored — that number comes from second-hand retelling and is not confirmed by the original report.
OpenAI’s response is that it invited the evaluation, that the findings helped strengthen safeguards as it scaled the system, that vulnerabilities affecting the production protection stack have been fixed, and that it will keep working with third-party evaluators. The evaluators did not independently verify the fixes.
The value of this kind of embedded assessment is turning “the lab says it is safe” into “an outside party has looked.” But how much gets disclosed determines how far it can be verified: with the report unpublished and the fixes unreviewed, what remains is self-report. In the same week, OpenAI and Uber rolled ChatGPT Work out to every employee, expanding how much production surface agent monitoring gaps now cover.
Sources:
- https://blog.redwoodresearch.org/p/red-teaming-openais-internal-agent
- https://x.com/MaxForAI/status/2113382989640212480
Theme 9: New benchmarks test what leaderboards miss
The day’s most counterintuitive item came from robotics evaluation. A new paper put GPT-6-Astra through 12 robotics benchmarks; it approaches saturation on the LIBERO suite and scored 81.9% on RoboDojo, but the paper concludes it cannot verify physical success — and one very long task burned about 9 million tokens and nearly two hours. High scores and tasks actually getting done are two different questions here.
The medical benchmark is more contested. Microsoft released the sequential diagnosis benchmark SDBench and the diagnostic system MAI-DxO, with the paper appearing in Science, and says it beats doctors fourfold on hard cases. That conclusion rests on a self-built benchmark and real case data; Microsoft AI chief Suleyman says medical AI has finally arrived, while outside media question the benchmark design and no third party has reproduced it.
The finance result reads more like a stress test. The Quant-Arena research benchmark gave nine frontier models historical data on 3,042 China A-shares and had them write code, build signals, backtest and revise strategies, then froze one version and tested it against sealed 2024 market data, including trading costs and simulated impact. Top scorer GPT-6 Astra returned 10.7% annualized while sitting in cash on 75% of trading days; only four of the nine models made money.
The voice track adds a detail dimension. Sierra’s τ-Entity benchmark tests whether voice agents can correctly collect, catch errors in and repair names, phone numbers, addresses and emails — 10 entity types, 200 tasks, six languages.
Another benchmark measures scale itself. A new framework built 500 million synthetic household personas from real survey data, using a compact decision model to simulate spending, cash and repayment paths; it reached 74.55% on consumer finance choices and has run across eight financial task families. The question is not how smart a single model is, but whether conclusions hold when agents are placed in an economy-scale simulation.
What these benchmarks share is breaking “can it be used” into small testable pieces. Vendor-built benchmarks still set the narrative, and the gap between self-testing and self-report can only be closed by outside reproduction.
Sources:
- https://microsoft.ai/news/building-medical-superintelligence/
- https://sierra.ai/blog/tau-entity-can-voice-agents-get-the-details-right
Theme 10: Listings and funding crowded into one day
Anthropic could file its S-1 publicly as soon as October 23. Yahoo Finance, citing three people familiar with the matter, said the company confidentially filed a draft with the SEC on June 1 and this would be the public disclosure; Reuters gave a tighter timetable: investor presentations starting the week of November 9, aiming to list in mid-November, with a potential valuation around $1.5 trillion, both subject to change.
A Newcomer long read adds the cap-table history: for years venture firms held back over governance structure and the safety-first strategy, and today no institution holds more than 2%; the company explicitly kept single-investor stakes at that level to avoid concentrated influence.
DeepSeek is on a different path. Bloomberg reported a new round of roughly 100 billion yuan, about $15 billion, with Tencent and CATL contributing the most; the figure was negotiated up from 50 billion yuan twice, and none of the three parties has publicly confirmed it.
Talent and products moved too. Warp Sciences, founded by former OpenAI chief product officer Kevin Weil, closed a $215 million Series A led by Lightspeed to synthesize and test molecules for drugmakers and speed up the most time-consuming experimental work in drug development. Sequoia partner Stephanie Zhan is preparing her own early-stage AI fund, with Sequoia and Jeff Dean as LPs, while keeping her Sequoia partnership. These items are media reports and company announcements; valuations and actual closing terms are unverified.
Sources:
- https://www.newcomer.co/p/the-story-of-anthropics-cap-table
- https://x.com/MaxForAI/status/2113310327131762737
High-value briefs
- iFlytek Spark X2.6: 293B total parameters, 30B active, 1M-token context; the company says it is the first model to train coding and agent capabilities entirely on domestic compute. Input is a limited-time 0.9 yuan per million tokens, and the 4B on-device version is free.
- Recursive Flint: an open coding model from Tian Yuandong’s team, about 40B parameters with a mixture-of-experts structure, scoring 89.7% on SWE-bench Pro v2; the team says Eureka Machine was involved in most of the training pipeline.
- TokenGate: a pluggable pruning module from Stanford and Berkeley researchers, inserted in shallow layers to pass only important top-K tokens upward; up to 1.4–11.3× faster time to first token with no accuracy loss, 1.5–3.4× lower peak VRAM, and training costs close to LoRA.
- SkyRL long-context training: on a B200 it pushed GLM-5.3-Flash to a 1M-token single-sequence training length at about 144.5 GiB peak, roughly 64× the out-of-the-box Megatron baseline, via sparse attention running on sparse kernels, a chunked and sharded indexer, and activation recomputation.
- OpenAI Decisions API: built on GPT-6 Luna for bounded problems such as classification, routing and scoring; the company says it is up to 10× faster than calling the same model through the Responses API.
- Replit Monitoring Agent: reads production logs and metrics to investigate errors and failed deploys, verifies fixes and reports by email when autofix is enabled, and can deploy fixes directly with auto-publish on.
- Kimi K3.1 and Qwen4 approach: anonymous models Kibitz and Kalmia appeared on Arena, and some infer from tokenizer characteristics that they point to Kimi and Qwen respectively, saying both are in online anonymous testing; not officially confirmed.
- X-VLA on APXInf: Tsinghua AIR’s 0.9B robotics model was ported to an inference engine by an agent in one day, with the six missing CUDA operators written by the agent — 8 commits, 5,300 lines of code, and one human decision. Maximum action error was 0.0032–0.0038, and CUDA Graph ran in 40.5 ms on a 4090 versus 93.4 ms for the original.
- PGX ephemeral databases: a PostgreSQL-compatible runtime in Rust where one process can hold hundreds to thousands of isolated databases, using sealed templates and a hot pool for millisecond creation, aimed at agents, CI and per-PR preview environments.
- Zijing Zhikang AI hospital: incubated at Tsinghua AIR and now in testing at multiple hospitals, giving each doctor a dedicated AI twin that can collaborate across departments; the company says its AI doctor exceeds 96% accuracy on MedQA and is updated through outpatient data fine-tuning.
- Chinese disclosures on RSI: in cases compiled by GovAI researchers, MiniMax had M2.7 analyze failure logs, plan changes, edit code, run evaluations and decide whether to keep or revert, running autonomously for over 100 rounds and gaining 30% on an internal evaluation set; Zhipu’s case focuses on self-optimizing inference infrastructure. Both improved the harness, not the weights.
- Google DeepMind lifetime non-disparagement: Neel Nanda says at least many UK DeepMind contracts contain a lifetime non-disparagement clause; the BBC reported that research engineer Sofia Liguori resigned over disagreement with the company expanding military partnerships and said publicly that in principle she is not allowed to criticize Google for life.
- Doubao pitching products in small talk: during the Double 11 shopping period it offered Douyin Mall coupons when the conversation turned to cold weather, and QuestMobile data shows monthly usage time fell from 144 to 133 minutes. Commercial pressure is testing the trust boundary.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 07:00 | Fireworks’ Ultrafast platform claims DeepSeek V4.1 Flash at 1521 tok/s | A single-post claim with no test conditions; speed competition is shifting from models to serving layers |
| 08:00 | Chrome DevTools adds an Application > Ads panel | Experimental CrUX ad metrics (including CPU and network usage) plus page-wide ad highlighting |
| 10:00 | Data marketplace HUD launched, listing private codebases and RL environments | Training data is starting to be priced per listing, and a new supplier role is appearing |
| 11:00 | A practitioner’s note on rewriting a goal criterion four or five times | Letting the main model judge goal completion, rather than a separate judge, is what lets long loops converge |
| 13:00 | Perplexity ships a Batch API on its Agent API, up to 50k requests per file | Batch work still runs the full agent loop, with search and fetch available |
| 13:00 | Anthropic security executive Jason Clinton: the past four days were bad, the next 18 months will be much worse | A frontline security judgment, more specific than a vague warning |
| 16:00 | eMudhra’s 8 intermediate certificate private keys were recoverable from public certificates, CVE-2026-111943 | Chrome removed emSign ECC Root CA - G3 from its root store; the root CA was unaffected |
| 17:00 | SemiAnalysis: Anthropic is A/B testing subscription quotas | The test group got about 20% more five-hour quota and about 20% less total monthly volume in API-equivalent terms |
| 19:00 | HeyGen Video 1 tops Artificial Analysis’s silent text-to-video leaderboard | It is built on MiniMax H3 with HeyGen post-training and ranks ahead of the base model |
Editorial conclusion
What matters most today is not how many more parameters some model gained. AI has leveled the speed of finding and exploiting vulnerabilities, agents are moving into accounts, payments and operations where consequences are real, and platforms are tightening parameters and defaults to take over risk judgment. Patches, authorization and monitoring are all half a step behind model capability.
Sources and method
Reviewed 20 hourly captures and 5 valid named sources from PT 2026-10-22, about 307KB in total. The signal pool was rich. Chrome Developers and XiaoHu.AI failed to fetch, and Google Research and the Claude Blog had no new posts that day. Vendor claims, single posts and unreproduced benchmarks are marked with their evidence boundaries in the text.
