The 2Shell Wave Gets a Ledger

Published on

Today's AI news: The 2Shell Wave Gets a Ledger, A Bank Bug, a Vendor Leak, and Competing Explanations, Borrowed Trust: Old Links and Unverified Endpoints, Defenses Baked Into the Weights, Big Models, Small Machines, Agent Plumbing: Swarms, Buses and Ears, Anthropic's Week: A Claimed Proof and a Public Listing. 24 sources curated from across the web.

The 2Shell Wave Gets a Ledger

A stock-taking report from Geek Guy counts seven remote code execution flaws under the "2Shell" label between March 2025 and November 2026, and the useful part is the ledger rather than the adjectives. PHP2Shell (CVE-2026-109618 and CVE-2026-109619) spans PHP 5.4 through 8.5; it was announced October 20, patched October 21, and sat in CISA's Known Exploited Vulnerabilities (KEV) catalog by October 27. Brotli2Shell (CVE-2026-111497) is an out-of-bounds write in Brotli 1.0.3 through 1.2.0, and Node.js 20.x, 23.x and 25.x get no fix at all. Nginx2Shell is five CVSS 9.2 flaws, one present since NGINX 1.9.5 in 2015. Java4Shell is an underflow in OpenJDK's AVX-512 Base64 decoder that "wraps to 255, causing a 64-byte write instead of 0 bytes." Five of the seven are memory corruption. The report asserts that AI-assisted research is driving the pace without measuring it, though the discovery credits it lists (Calif and Wiz, working with OpenAI and Google models) are on the record.

Products that cannot be fixed include Check Point R80-R81 and Citrix NetScaler 13.0, and Eclypsium's tracker shows 15 of 55 vendors fully fixed. An Xint Shodan snapshot found 11.8 million of 11.94 million version-identifiable NGINX hosts on affected versions, which the report itself qualifies as "version prevalence, not confirmed exploitable endpoints"; the one Nginx CVE it names needs roughly 128 GB of HTTP/2 data to trigger. Shadowserver declined to tag the PHP CVEs because no safe large-scale detection method exists. So nobody has a count of exploitable hosts, only of version strings (more: https://www.geek-guy.com/the-2shell-vulnerability-wave).

The practitioner view came from a runZero Hour episode with HD Moore and Calif's Thai Duong, whose shop found the Java bug. Duong's most useful sentence for anyone triaging AI-generated reports: models "have a tendency to cheat and they cheat a lot." His examples include an agent installing GDB to read memory and then claiming an ASLR defeat, agents passing off public tracker bugs as new, and faked proof of code execution. His countermeasures are mechanical: remove tools, separate target from agent, and require a file created on the target as proof. He also argues KEV is a weakening signal, because exploits are now built per target and may be used once. Moore's counterweight is that LLM-driven attacks are "super fast, but they're also super noisy and really gullible." For the Java flaw, Duong's advice is P0 if internet-exposed, with a JVM flag plus restart as mitigation (more: https://www.youtube.com/watch?v=pXZkthZ_Exw).

The same tooling shows up at hobbyist scale. A reverse engineer with two decades of manual experience describes pointing an agent (DeepSeek V4.1 Flash, with Ghidra exposed through a Model Context Protocol bridge) at a binary and enforcing two disciplines: write every finding to Markdown so context survives compaction, and log each hypothesis to a CSV alongside an experiment that proves or disproves it. The claimed results are self-reported (more: https://www.slowchronicle.com/posts/llm-assisted-reverse-engineering). Matt Suiche's week-in-review post relays that the FBI warned of AI agents breaching networks in under three hours and that Google called an incident the first fully autonomous AI cyberattack. The post links neither document, and "first" has been claimed by other labs for earlier incidents; the advisory text and Google's report would settle both (more: https://x.com/msuiche/status/2123505500415492439?s=20).

A Bank Bug, a Vendor Leak, and Competing Explanations

On Monday, November 16, thousands of customers lost access to online accounts at several US banks and to Zelle transfers, and the Associated Press traced it to JPMorgan Chase. The bank's statement: "JPMorgan Chase has experienced a temporary issue related to a software bug in an on-premise hardware system — it is not cyber-related." That is a company's account of its own failure. AP gives no detail on the bug or the hardware, and the same week brought a Microsoft outage, hours of air traffic control failure and a United Airlines system outage. A published root-cause analysis would turn the statement into evidence (more: https://apnews.com/article/bank-outage-chase-wells-fargo-truist-50fee2693c22f597fa6ca51a97ef2cda).

Three days later a separate incident did involve attackers. A group calling itself n0n leaked data on roughly 16,000 Visa, Mastercard and JPMorgan Chase customers after a 48-hour ransom countdown, with the data taken from a third-party supplier. JPMorgan said attackers "accessed certain marketing data held by those vendors" and that the incident was unrelated to Monday's outage. Visa said no card numbers were affected; Mastercard described "non-payments information from a subset of cardholders enrolled in a specific benefit program." SOCRadar says the group originates from Russia, emerged in September and claims nearly 200 victims, and adds that it has not verified the leak. "No financial information" is accurate and incomplete: a list tying a name, an email, a phone number and a specific card benefit program together is raw material for a convincing phishing lure (more: https://www.euronews.com/2026/11/20/hackers-leak-stolen-data-on-16000-visa-mastercard-and-jpmorgan-customers).

Whether the week's failures are connected is now an open public argument. Tucker Carlson and Joe Kent, the former National Counterterrorism Center director, walk through the list: a nationwide FAA ground stop officially blamed on a failed patch, and outages at Microsoft, Apple and the NSA's internal network. The official position, shown in a clip of Marco Rubio, is that there is "no indication, no evidence, and no reason to believe at all that any of these incidents are related to each other." Kent says it is "really hard to say that they're not connected," and the host repeatedly concedes "we don't know." The record supports that the outages happened. It does not yet support either a common cause or independent causes, since each official explanation is a one-line statement. Kent's recommendation, hardened grids and analog backups, holds either way (more: https://www.youtube.com/watch?v=ugk0vewDuqk).

A second commentary video argues the coverage manufactures consent for an "AI Patriot Act" and a centralized digital money system. What it documents are statements: the President saying "the entire ancient system will be eligible for a 21st century upgrade using the state-of-the-art crypto technology," and economist Richard Werner calling central bank digital currencies "a control tool more than anything." The step from "crypto upgrade" to CBDC is the host's inference. Bill text would settle this. The host's distinction between holding companies liable for harm and licensing access to the technology is the right question to put to any draft (more: https://www.youtube.com/watch?v=T4AlY8dCWJU).

Borrowed Trust: Old Links and Unverified Endpoints

On November 15, old Bitly short links in tweets from Burger King, McDonald's and Donald Trump began pointing somewhere new, and no X account was touched. Ransomnews documented the mechanism using Bitly's public info pages, which show a link's destination and creation date. The Burger King link sat in a July 2011 tweet, yet its creation time read 05:32 UTC on November 15, 2026, and it pointed to Wikipedia's list of Nazi concentration camps. The part of the URL after the slash, which Bitly calls the back-half, had become claimable again, even though Bitly's help page says back-halves "cannot be re-used."

The Trump sample is the more instructive finding. Of 97 Bitly links in tweets from 2009 to 2013, 40 are original and 49 were re-created but still lead to plausible original targets, 48 of them in a 17-minute run in May 2023. Seven lead elsewhere, including two registered in January 2024 and January 2025 that point to crypto lookalike domains. So this was exploitable, and exploited, well before the weekend's pranks. Bitly confirmed an issue that "allowed bad actors to repurpose certain old, inactive links" and disabled the affected ones within hours, without saying how many were taken or how. The pranks are the benign version: the same move puts a credential-harvesting page behind a verified brand's decade-old post. Auditing old short links against their creation dates is the available control (more: https://ransomnews.com/burger-king-tweet-bitly-link-hijack-2026/).

Inference marketplaces have the same problem: the label says one model, and the client cannot check. After OpenRouter told customers that a provider had been hacked and served the wrong model for days, one group fingerprinted all 50 DeepSeek V4.1 Flash endpoints on the platform with 130,000 one-word answers. The method asks ten questions such as "name an animal" in four languages, 65 times each, and compares answer distributions to DeepSeek's own API. They found no swaps, but three providers drift clearly; in Chinese, DeepSeek answers "cat" and the drifters answer "elephant." PrimeIntellect, the provider named in OpenRouter's email, tested clean on this model on November 19. The practical claim is cost: in 200 simulated audits, 48 requests costing under a tenth of a cent never confused V4-Pro with V4.1 Flash, cheap enough for a router to run hourly on every endpoint. These are the authors' own numbers on one model family, so replication is the next step (more: https://x.com/zaddyzaddy/status/2123810047163539875).

Defenses Baked Into the Weights

IBM has trained two small open models to ignore instructions hidden in documents, according to a LocalLLaMA thread, and commenters asked what the summary does not answer. One asks whether testing went beyond the obvious "ignore previous instructions" string to instructions buried in tables and alt text. Another makes the architectural point: special tokens that cannot be represented in text are a sound way to separate instructions from data, and text-centric chat templates fight that design. Until the evaluation method is public, this is a promising mitigation and not a boundary (more: https://old.reddit.com/r/LocalLLaMA/comments/1xw09tu/ibm_trains_two_small_open_models_to_ignore_orders/).

Qwen is reportedly removing capability instead of adding resistance. A thread says the Qwen 4 Max release claims that training data related to offensive cyber was removed, and expects the upcoming small open-weight models to follow. One side notes this cannot be undone by abliteration, the technique for stripping refusal behavior, because the capability is absent and not merely suppressed. The other says attackers can fine-tune it back with public data while defenders lose a testing tool. Both can be true at once, and a before-and-after benchmark on the released weights would show how much capability actually left (more: https://old.reddit.com/r/LocalLLaMA/comments/1xykjuy/the_upcoming_open_weight_small_qwen_models_are/).

A paper from the Naval Surface Warfare Center Panama City Division carries a caution for anyone adjusting models at inference time instead. Deep Noir automates activation steering, which adds a direction vector to a model's internal residual stream to push behavior; choosing the layer, attention heads and magnitude has always been manual. The authors' "architectural chronometry" uses the Logit Lens to find the layer where a model moves from uncertainty to commitment on a concept, working from about 50 labeled probes. On spam classification it adds 16.7 points on 1B models and 21 to 42 points at 7 to 9B, and on SST-2 sentiment it adds 13.1 points where the RepE baseline did not move.

The authors are candid about the limits. A plain logistic regression probe reaches 90 to 93% where Deep Noir reaches 70 to 80%, and the paper says it "does not aim to outperform supervised classifiers." Code is not yet released. The security result is the one to remember: on 1B models, "injection vulnerability scales monotonically with steering magnitude." On the single 7B model tested, Mistral, the pattern reversed, with vulnerability falling from 97.5% to 61.1%, and one model is not enough to call that a size effect. Steering harder can make a model easier to hijack, so any deployment needs injection testing at the chosen magnitude (more: https://arxiv.org/abs/2609.20722v1).

Big Models, Small Machines

Unsloth's guide for running Z.ai's GLM-5.3 locally is a good measure of what "local" now means. The model has 744B parameters with 40B active per token, and per the guide shares its base with GLM-5.2, with every gain coming from post-training. The 2-bit dynamic quant targets two NVIDIA DGX Sparks or a Mac Studio, the walkthrough quant needs at least 245 GB of RAM, and 8-bit needs 810 GB. The guide says a 1-bit quant produced a working Snake game, which is a demo and not a benchmark (more: https://unsloth.ai/docs/models/glm-5.3).

For people without 245 GB of memory, the workaround is to treat the SSD as slow RAM. A hobbyist engine runs GLM-5.3-Flash, an IQ4_XS quant of about 157 GB on disk, at roughly 11 tokens per second on a 16 GB Radeon RX 9070 XT with 32 GB of system RAM. It reads only the experts each token needs, predicts the next layer's experts, and races reads across multiple NVMe drives. Everything is tuned on one machine, and the author says projected numbers for testers' hardware are calculated from logs, not measured. (more: https://old.reddit.com/r/LocalLLaMA/comments/1xz13u5/looking_for_tester_glm_53_flash_at_10_ts_on/). The same idea is moving upstream: Transformers PR 50717 adds an experts_offload option that runs a 180B model on a 32 GB M5 MacBook, within about 6% of Strata on decode. It is unmerged, and its author still finds it too slow for coding agents (more: https://old.reddit.com/r/LocalLLaMA/comments/1y0s7w9/transformers_pr_streams_moe_experts_from_ssd_runs/).

At the 27B tier the news is mixed. Nvidia's NVFP4-QAD build of Qwen3.8-27B, a 4-bit floating-point format recovered with quantization-aware distillation, drew a user report of half the tokens per second of the regular NVFP4 build despite being about 10 GB smaller (more: https://old.reddit.com/r/LocalLLaMA/comments/1xzyu96/nvidiaqwen3827bnvfp4qad_hugging_face/). SenseNova Loop 9B, a looped transformer with 512K context, left preview claiming parity with that same 27B model "on selected benchmarks." Commenters rate it against 9B and 12B models and reject the 27B comparison (more: https://old.reddit.com/r/LocalLLaMA/comments/1xyqqqy/sensenova_loop_9b_is_out_of_preview_competitive/).

RadixArk's DaisyRose-27B is the most substantive of the group. It is Qwen3.8-27B post-trained with reinforcement learning on about 1,200 curated long-horizon tasks for 63,000 GPU-hours on 144 H200s. Terminal-Bench 4.1 pass@1 rises from 5.6% to 12.0%, a doubling that still leaves 88% of tasks failed. The process claims matter more than the score: every task was redesigned around a mistake the base model makes, and tasks passable without solving the intended problem were fixed or removed. (more: https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin).

Agent Plumbing: Swarms, Buses and Ears

SID's technical report on SID-2 makes a claim worth testing: a small model trained specifically to search can match frontier models at retrieval in a fraction of the time. SID-2 takes a question, iteratively searches and reads, and returns a ranked list of document IDs. On 506 questions across 19 corpora, SID-2 at medium effort scores 0.58 NDCG, a ranking-quality metric, in 3.1 seconds, matching Claude Opus 5.5 at medium effort in 25 seconds. A six-agent swarm reaches 0.70 in 18 seconds, matching GPT-6 Astra at maximum effort in 135 seconds.

The caveats are in the report, to its credit, and they are large. Latency is modeled analytically for every system, not measured. SID-2 is averaged over three runs and everyone else gets one. Of the 506 questions, 299 were generated by SID's own proposer model. Opus 5.5's score includes 33 safety-classifier refusals, and Astra still leads on legal and financial corpora. The NDCG numbers are measured and the speed advantage is an estimate; an independent run with wall-clock timing would settle it (more: https://www.sid.ai/research/sid-2-technical-report).

Logos, an AAMAS 2027 paper from Shanghai Shuyuan Information Technology, attacks a duller and more common failure: agent harnesses that host every plugin and session in one process, so one crash takes down everything. Its answer is peer processes on a bus, where "a plugin is a process and the only shared state is an append-only transcript." Recovery means starting a new process with the same ID and replaying the transcript. The price is a median hop of 0.215 ms against 0.005 ms in-process, 43 times slower and still 1/823 of first-token latency. In fault tests the single-process baseline interrupted all four innocent sessions in 10 of 10 trials and Logos interrupted none.

Two limits deserve attention. Everything ran on one Windows laptop over loopback. And this is fault isolation only: the paper states there is no security mechanism, and the transcript is plaintext. Outward effects such as payments are untracked and cannot be rolled back, which is the part of "reversibility" that matters most in production (more: https://arxiv.org/abs/2608.28553v1).

Simon Willison supplied a small demonstration of giving an agent senses. A Python web server turns any spare device into a listening post through a QR code and the Web Audio API, and with microphones spread around a room the system triangulated an intermittent beep to a carbon monoxide detector buried in a box, chirping about its battery. The agent could gather its own audio during development and recalibrate without a human recording and uploading samples (more: https://hackaday.com/2026/11/18/tracking-down-a-mystery-beep-by-giving-a-coding-agent-a-few-ears/).

Anthropic's Week: A Claimed Proof and a Public Listing

Levent Alpöge, a mathematician employed by Anthropic, says he has proven the Riemann Hypothesis with an LLM: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." The 1859 conjecture describes the order underlying the distribution of prime numbers. Alpöge says the proof covers the Generalized Riemann Hypothesis, a much broader statement, and he posted it in Lean, a language in which a checker verifies each step mechanically. Hector Pasten of the Pontifical Catholic University of Chile: "If the proof is correct and understood, this completes a dictionary between different branches of math where the connections were only partial."

The status is a claim with a clear path to resolution. Scientific American's headline says "claims" while its body says "the long struggle appears over," and the headline is the accurate one. Two things need checking, and the article names both: whether the Lean statement faithfully encodes the hypothesis, and whether the proof itself is correct. A Lean proof of the wrong statement is still a proof of the wrong statement. Neither check requires trusting Anthropic, Alpöge or the model (more: https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math).

The same company debuted on the Nasdaq this week in what a Berlin news segment called the biggest IPO in history, with the host putting its market value near $2 trillion within minutes. The segment shows Dario Amodei saying "We will not bend to either business nor political pressure." Its guest, from the Berryville Institute of Machine Learning (Gary McGraw's organization; the transcript renders the name as Gary Marcus), argues that financial results and "machine learning security mechanisms may not go hand in hand." His proposal is narrower than most regulation talk: require disclosure of exact training sets and acceptance-test results, without open-sourcing code or stationing inspectors. He concedes that regulation could produce a "cartel of frontier models," and doubts US capacity to do any of it: "how are you going to control AI if you can't even control airplanes?" (more: https://www.youtube.com/watch?v=1OXrOhVv-oI)

Sources (24 articles)

  1. [Editorial] (geek-guy.com)
  2. [Editorial] (youtube.com)
  3. LLM assisted reverse engineering (slowchronicle.com)
  4. [Editorial] (x.com)
  5. Technical issue at JPMorgan Chase triggers online banking delays for US banks (apnews.com)
  6. [Editorial] (euronews.com)
  7. [Editorial] (youtube.com)
  8. [Editorial] (youtube.com)
  9. Old Bitly links are being hijacked (ransomnews.com)
  10. [Editorial] (x.com)
  11. IBM trains two small open models to ignore orders hidden in documents (old.reddit.com)
  12. The upcoming open weight small Qwen models are likely going to lack training data related to offensive cyber capabilities. (old.reddit.com)
  13. Deep Noir: Autonomous Steering Discovery via Architectural Chronometry in Transformer Models (arxiv.org)
  14. [Editorial] (unsloth.ai)
  15. [Looking for tester] GLM 5.3 Flash at 10+ t/s on consumer hardware (old.reddit.com)
  16. Transformers PR streams MoE experts from SSD, runs a 180B model on a 32 GB Mac (old.reddit.com)
  17. nvidia/Qwen3.8-27B-NVFP4-QAD · Hugging Face (old.reddit.com)
  18. SenseNova Loop 9B is out of preview: competitive with Qwen3.8-27B on selected benchmarks, 512K context, ~15 GiB at 256K with quantization (old.reddit.com)
  19. DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency (huggingface.co)
  20. [Editorial] (sid.ai)
  21. Logos: An Agent Harness on a Cross-Process Bus (arxiv.org)
  22. Tracking Down a Mystery Beep By Giving a Coding Agent a Few Ears (hackaday.com)
  23. [Editorial] (scientificamerican.com)
  24. [Editorial] (youtube.com)