AI router privacy: 2 million canaries expose the leaks

Two million AI canaries found prompts leaking through resellers. Signal blocks AI agents from screens, and a Claude-assisted Riemann claim awaits verification.

Top Stories

Two million canaries show where your prompts leak

Ars Technica reports on a five-month experiment by HD Moore, chief research officer at security firm Dragos, who assembled a set of AWS access keys tied to empty accounts, SSH passwords, unique URLs, DNS names and references to private git repositories, then put each secret into exactly one prompt sent to one provider. Nearly two million prompts went out to AI providers. “We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced,” Moore wrote in Dragos’s report, adding that the team found “security breaches, prompt theft, and widespread failures to protect customer data.”

The routing is the part that matters. Many apps do not call Anthropic or OpenAI directly. They hire a router or gateway, which reads the plaintext of the prompt to decide which inference provider is cheapest or best suited, and then the prompt can be forwarded to another router or to a pool of models running inside different hosting providers. Every hop in that chain is a place where the prompt can be logged. Moore traced roughly 2,000 inference routes.

Three hosting providers were breached. In the largest incident an attacker reused “a couple hundred” of the AWS keys, and the requests carrying them had all gone out in a few short windows in late October. Moore’s writeup notes that some of the stolen keys came from requests the provider rejected before a model ever ran, which tells the researchers the attacker was copying requests “at the provider’s front door, and not somewhere deeper in the inference stack.” The attackers listed S3 buckets, searched secrets managers and parameter stores, enabled models in AWS Bedrock and added their own IAM users. The stolen bait keys were used “tens of thousands” of times in AWS API calls. In two of the breaches some stolen requests were marked Zero Data Retention, and Moore’s verdict is blunt: “A privacy flag doesn’t help when the attacker is already inside.”

Resellers themselves turned out to be a leak point. Moore identified them by sending the same model through different sellers at low prices with request logging on, then watching for credentials being reused through one seller and not another. One seller received two prompts and a person on home broadband opened both secret URLs about a day and a half later. The most serious finding was a scraper. Moore wrote that a Chinese AI lab’s systems “fetched the secret URLs from more than half of the prompts that we sent to its models between July and October, including roughly two-thirds of the requests marked as ‘no data collection’ or Zero Data Retention,” with the scraper typically showing up “two or three days after the request.” Several US hosting providers resell that lab’s models, so picking a US provider did not stop it, and a git client on a different cloud later tried to clone the private repositories named in those prompts.

Moore’s conclusion matches what self-hosting advocates have argued for years, though now with two million data points behind it. He says the most effective protection is hosting models on premises, and that the next best option for anyone who cannot afford that is to receive service directly from a first-tier provider that provides “one contract and a privacy policy to read.” Users should also avoid default provider or reseller pools and pick upstream providers themselves. Dragos said it was working with about a dozen providers, one of which has published a preliminary security notice and another of which has not publicly disclosed its breach.

Signal blocks AI agents from reading your screen, and calls it an existential threat

Signal has restricted screenshot capture inside its app on iOS 27 and on Android so that operating system agents cannot read message content, Signal president Meredith Whittaker writes that these agentic capabilities are undermining the app’s ability to guarantee privacy. The Verge quotes her saying the agentic features built into operating systems are “undermining Signal’s ability to independently guarantee privacy.” On iOS 27 Signal blocks screenshots in the app, and on Android it detects when an AI assistant tries to read the screen and switches on “Screen security,” covering Google’s Gemini Intelligence and Apple’s Siri AI. Users who want screenshots back can disable the protection in Signal’s app settings, but doing so will “potentially expose your (and your contacts’) Signal content to these AI agents, and break our privacy guarantees.”

The fix does not hold everywhere. Whittaker admits the remedy is “incomplete,” because it does not cover agents that reach app data through the device file system. Signal cannot block screenshots on macOS at all, and says plainly that “we cannot reliably block screenshots on macOS,” pointing macOS 27 users instead to the excluded apps section in System Settings, Siri. On some non-Pixel Android builds Signal also says it cannot reliably detect an attempted screenshot when screenshots are already switched off. Signal is asking Apple, Google and Microsoft for “simple technical tools and accurate documentation” so app developers can restrict what agents can reach without wrecking the user experience.

Third-party agents are the other half of the problem. Signal singles out agents such as Meta’s Muse for using broad accessibility grants as “a skeleton key that provides access to AI agents.” The wider framing from Signal is that the current integration of agents into operating systems, without those remedies, is “an existential threat to privacy at the application layer,” and that “Signal is the canary in the coal mine, but this affects the entire ecosystem and should galvanize anyone who cares about these fundamental rights.” The Verge reports that Signal used a similar tactic last year against Microsoft’s Recall feature.

METR’s capability update: 3.6-month doubling and a measurement problem

METR’s quarterly capability update puts the headline number at 3.61 months, with a 95% confidence interval of 2.08 to 4.94 months, for the time horizon doubling rate on its combined TH1.1 and TH2 suites. That replaces the earlier TH1.1-only estimate of 105 days. On the narrower verifiable subset of TH2, METR estimates a doubling time of 1.65 months, with a 95% confidence interval of 0.79 to 3.99 months, and it says the cause of the gap is uncertain. Current frontier models reach an estimated time horizon of up to 27 hours on that verifiable subset. On the combined suites, METR reports 50% time horizons of 40 hours for GPT-6 Astra and 39 hours for Opus 5.5, both with very wide intervals, and much tighter ones once the estimates are adjusted.

The measurement side is where METR is more cautious, and that is the part worth reading. METR found that current frontier models beat the human baseliners it recruited on all three tasks where it attempted to acquire baselines, and establishing those human baselines cost roughly $90,000 at $200 to $350 an hour across seven attempts. On the Budget NanoGPT speedrun, the best agent run reached a 15.7% speedup in 48 hours while the strongest human contributor reached 41% in about 60 hours. And on the task of designing evaluations, METR reports that “None of the tasks agents produced were sufficiently useful to us to warrant inclusion in this post.” AI judges overestimate an output’s quality when the output comes from a similarly strong or stronger model, and Opus 5.5, the judge with the highest agreement with METR’s human scorers, rates Astra’s tasks on average about 0.2 out of 1 more highly than METR’s humans do.

MIT Technology Review explains why the original plot, which showed task duration doubling roughly every four months, spread so widely and got read for more than it could support. The updated plot adds more tasks, some of them estimated to take humans far longer than those in the original suite, and in the new report the human completion times are mostly estimates rather than measurements. Verifiable tasks such as coding and math can be scored automatically, while judgment-heavy work like choosing a research direction is much harder to quantify. Sayash Kapoor said “We’re at the point where AI evaluations need to get more creative,” and Matt Putz said “By their nature, hard-to-verify or judgment-laden tasks are much harder to study, and especially to quantify.”

Two claims on the Riemann hypothesis, one made with Claude and one without AI

Levent Alpöge, an Anthropic employee, says he proved the Riemann hypothesis working with Claude, and The Verge reports that his solution is presented in Lean and has so far appeared only on his own website. His statement reads: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis. In fact, the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis, and I expect, with more time, the Extended Riemann Hypothesis too.” Simon Willison quotes the same passage and links to Alpöge’s own document, in which he says Claude “far outpaced me near the finish line.”

The claim is not yet verified independently. MIT mathematician Andrew Sutherland successfully compiled the Lean proof, TechCrunch reports, while most mathematicians in the discussion appear confident it would hold. Alpöge texted Anthropic CEO Dario Amodei about the solution “late Sunday night” and suggests he held off publishing until after the company’s IPO on Thursday. Anthropic has not commented, and in August the company said it was working on the Riemann hypothesis with Claude while noting at the time: “We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.”

Running the other way, Stanford postdoc Alessandro Masullo released a proof of the Generalized Riemann hypothesis that he says took twelve years, and it is not written in Lean. Ravi Vakil, a Stanford professor and the president of the American Mathematical Association, posted it and stressed its provenance: “it is all human, with no AI.” Masullo introduced his proof by telling anyone discouraged by “the pace and tone of the current climate” that “we are still here. We are very much alive. We are not going anywhere.” The Clay Mathematics Institute administers the Millennium Prize and pays $1 million for each problem solved. Under its rules a proposed solution must be published and withstand at least two years of scrutiny, during which time it must have “received general acceptance in the global mathematics community.” Only the Poincaré conjecture has been officially solved since the prize was established.

Inkling 2 Mini: a 20GB open-weight multimodal model built for local setups

Thinking Machines Lab has released Inkling 2 Mini, a multimodal mixture-of-experts model with about 26B physical parameters and 3B active per token, a 1M-token context window, and Apache 2.0 weights. The Hugging Face announcement describes a looped design with 24 stored decoder layers where the middle 12 run three times, so the model card lists 26.4B total and 3.4B active parameters, and notes that “Since the model runs with three passes, there are 5.4B effective active parameters per token.”

The checkpoint is the practical part. The model card says it is “distributed as a single checkpoint of about 20GB, in mixed NVFP4 and MXFP8 precisions,” and that local runs need “Apple silicon devices with 32GB or more memory (MacBooks, Mac minis and Studios)” or a GPU with 32GB or more of VRAM recommended, with one NVIDIA DGX Spark also listed as supported. llama.cpp builds are published separately in the ggml-org GGUF repository, where the IQ4_NL quant file is 17.63GB and the Q8_0 file is 27.82GB, alongside a 52.23GB BF16 build and separate multimodal projector files.

On quality, the announcement says Inkling 2 Mini “leads in 32 benchmarks out of 48” against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL, and calls it “state-of-the-art in its class” on the Artificial Analysis Intelligence Index v5. Against Qwen3.8-27B specifically it ties on Terminal-Bench 4.0, trails on AutomationBench-AA v1.0.6 and leads on FrontierCode v2.0. Support covers Transformers, llama.cpp, SGLang and MLX, with Ollama and Exo Labs on Apple silicon also listed.

House committees say Chinese open-weight models are already inside US companies

The House Select Committee on the CCP and the House Committee on Homeland Security have published a joint investigation report arguing that Chinese open-weight models are now widespread across the American economy and that their use is difficult to track. The committees say adoption is driven by their “inexpensive, customizable, and competitive” characteristics, that the US has no similarly competitive open-weight option at the required price and performance, and that Chinese developers used “proxy networks and fraudulent accounts” to collect outputs from American frontier models.

The military details are specific. Personnel associated with the PLA Cyberspace Force visited Zhipu’s Beijing headquarters multiple times, most recently in August 2026. A July 2026 PLA procurement for a large cyber range required integrating DeepSeek to support “intelligent attack” and “intelligent penetration.” Another command information system was built around Alibaba’s Qwen and DeepSeek, and a separate 2026 military project required DeepSeek, Alibaba, Moonshot and MiniMax models for an internet-isolated network used in military AI-agent development.

The committees have pressed American companies for information. Anysphere and Airbnb received initial investigation letters in April 2026 and have since engaged with the committees, and DoorDash received a similar information request in July 2026.

OpenAI says it cut reasoning-theft attacks from 54 percent to 0.2 percent

OpenAI has published its approach to detecting and disrupting adversarial distillation, the practice of extracting a model’s protected reasoning traces by attacking the API. The company says the method combines adversarial testing, model training, classifiers and API-level protections, with input classifiers that flag extraction attempts, output classifiers that catch protected reasoning in assistant messages or tool-call arguments, and API controls that bind protected reasoning to the conversation that produced it.

The reported numbers come from OpenAI’s own adaptive black-box red-teaming across six attacker models. The average attack success rate for extracting protected reasoning from an earlier request fell from 54.0% against GPT-6 Astra with the original safeguards to 0.2% against GPT-6.2 Astra with the updated ones, and from 22.9% to 0.9% for the current request. Weekly enforcement against unauthorized campaigns ran from “thousands to tens of thousands of accounts.” The safety framing is that a distilled model inherits none of the original safeguards, so it can help with work the source model would refuse, in areas like cyber and biology. These are vendor-reported figures with no independent verification.

Trump threatens a “tremendous tax” on data centers built outside the US

President Donald Trump said his administration will put a heavy tax on companies that build data centers outside the country, The Verge reports, speaking at a press conference on Friday after his summit with China President Xi Jinping and Russia President Vladimir Putin. “We’re not going to allow our companies to build data centers outside of our country,” Trump said, and warned that “China will take them all.” He added: “If they do that, they can, but we’re going to put a tremendous tax on them.”

The detail is thin, and Trump said so. He said “I don’t know if that’s a breaking story. But I haven’t even said that to my people,” flanked by Vice President JD Vance and Secretary of State Marco Rubio, and he shared no further plan. The threat arrives while states and cities are restricting new data center applications, which the Verge notes could push builders abroad. It also runs into a new Democratic majority Congress being sworn in within a couple of months.

EFF and 25 other groups want Section 702 out of a must-pass bill

Section 702 of FISA expired in June 2026 after Congress failed to reach agreement on reauthorizing it, and the statute lets the government compel telecommunications companies to hand data to the NSA for overseas national-security surveillance. EFF and 25 other civil liberties groups are asking lawmakers not to revive it by attaching it to the National Defense Authorization Act, a continuing resolution, or a rushed year-end funding bill meant to avert a shutdown.

EFF’s objection is procedural as much as substantive. The group’s post calls the maneuver “both undemocratic and suppress any real chance at reform,” and the joint letter warns that using the NDAA would “circumvent or to cut short this long-running bipartisan debate” over civil liberties and constitutional rights. Adding language that has not been properly debated “would be a serious procedural and substantive failure, and a betrayal of the democratic process,” and EFF’s conclusion is a direct plea: “We once again urge Congress not to put Section 702’s future in a must-pass bill like the NDAA.”

The stakes are in the access the statute creates. EFF says that because the world is globalised, communications between people in other countries and in the United States often get hoovered up, and that once they are, the FBI and other agencies can then search US persons’ data without a warrant.

Honey traps now target almost everyone working in AI

Security researchers are warning that honey-trap operations against AI workers have spread well past founders and chief executives, TechCrunch reports, describing approaches that arrive as investors, customers, recruiters or consultants, build a friendship or romance, and then extract roadmaps, customer lists or access to technology. The Wall Street Journal reported that OpenAI has begun warning employees about outsiders trying to get close to them in pursuit of company secrets, and that anyone with access to sensitive information, from board members to executive assistants, can be the target.

Former senior CIA officer Aaron Brown, who founded Lumbra, says a spy “wants what the CEO can touch.” He describes the danger as real rather than performed: “It doesn’t feel like manipulation. It feels genuine, even lucky.” His warning for company leaders is short: “Recognizing you’re a target does [help].” KnowBe4’s chief deception strategist Perry Carpenter argues that attackers skip technical defenses: “Why fight through a firewall when you can get close to the person who can open doors, influence decisions, or make requests nobody wants to question?” He recommends watching for unusually frequent coincidences, persistent curiosity about work and colleagues, requests for secrecy, and presses for introductions, and urges firms to “eliminate executive exceptionalism,” warning that “One of the most dangerous security cultures is one where nobody feels comfortable questioning the CEO.”

The reason this got easier is the same technology. AI can research a target, build fake profiles and fake backgrounds, run fake voice calls and hold up a whole invented history, which is what makes the scheme cheap to plan and scale. Carpenter is careful about what that does and does not mean, saying “the deception ultimately plays out through human relationships.”

Quick Hits

  • A LoRA adapter trained with reinforcement learning inside a coding harness took Qwen3.8-27B from 39% to 55% pass@1 on Terminal-Bench 2.1 while cutting average token use from 2.9K to 1.5K per task. Specialist adapters for pi, Hermes Agent and OpenCode merged into one super-adapter, and the GGUFs are published. Hugging Face
  • Chips have become data center loan collateral, and the collateral is depreciating. Some H100s are selling for about half their original cost basis from 2023, according to Compute Credit Index Research, while lenders behind SpaceX’s $40 billion chip financing arrangement asked that none of the purchased chips be used for orbital data center plans. TechCrunch
  • Canada’s Bolo Program unveiled AI-generated lifelike images of the country’s 25 most wanted fugitives at a Toronto news conference. Executive director Max Langlois said “If it feels uncomfortable, that’s the point,” and the program is offering $250,000 for information leading to the arrest of the top suspect, Satinderjeet Singh. CBC News
  • The Justice Department is using the AI-writing detector Pangram against defense attorneys in a Texas voter fraud case, with DOJ lawyers writing that “This case has turned into defense by ChatGPT” and describing opposing motions as AI slop. 404 Media
  • A small business owner and his Meta Muse agent are suing Meta after Meta’s moderation disabled the Facebook pages the agent ran, with the complaint seeking $75,000 in damages. 404 Media
  • 404 Media profiled the person behind the AI-powered fake ID site OnlyFake, alongside reporting on the Muse lawsuit and the Justice Department’s use of detection tools. 404 Media

Worth Watching

  • Whether the Riemann claim survives review. The Lean proof has been compiled by at least one outside mathematician, but the Clay Mathematics Institute requires publication, at least two years of scrutiny and general acceptance in the global mathematics community before the prize is considered. The separate human-only Generalized Riemann proof is not in Lean and is much harder to check.
  • Whether the router-layer breaches get disclosed. Dragos said it was working with about a dozen providers, one of which has published a preliminary security notice while another has not publicly disclosed its breach. The three hosting providers involved in the AWS key theft are the ones that matter most.
  • Whether METR can keep measuring what it measures. The gap between the combined-suite doubling time and the faster verifiable-subset rate is unexplained, and METR says models have not yet produced an evaluation worth including in its study. As tasks get harder to score automatically, the chart may become harder to read before it becomes easier.
  • Which bill carries Section 702. The NDAA, a continuing resolution and a year-end funding bill are all on the table, and each would let the surveillance authority pass without the standalone debate EFF and its 25 allies asked for.
  • How the data center tax threat meets the state-level moratoria. Local restrictions are already pushing builders toward other countries, and a federal penalty aimed at that flight is difficult to design without details. Watch for whether an actual proposal replaces the press conference.