Ground Truth.
AI, checked against the source.

AI News Today, Nov 16: AI in Detail — FBI Warns of ARTEX AI-Assisted Intrusions

2026-11-16 · Breach Protocol: Inside the AI Blackbox — full transcript

The FBI warns of ARTEX in real intrusions; Anthropic changes cyber safeguards; Alibaba and Google release model tools; a candidate physics construction, an OpenAI publisher deal, a game-playing agent and a narrow hiring-post statistic. This audio was created with Boson AI's Higgs Audio -- https://www.boson.ai/higgs-audio

Listen on Spotify · Watch on YouTube · Pocket Casts

ARTEX used in real intrusions

Eris: <|prosody:expressive_high|>The FBI says attackers have used an AI-assisted platform called ARTEX in real intrusions involving South Korean financial organizations and a cloud backup provider. The agency's advisory describes software that connects language models to specialized agents that can operate attack tools. How much of that is the model itself?

Vestra: <|prosody:expressive_high|>ARTEX doesn't contain its own language model. The FBI describes it turning a broad instruction into jobs for agents that perform tasks such as discovering accounts, abusing credentials and collecting data. The model is one part of a larger system with tools and permissions.

Eris: In the backup incident, Microsoft says the group it calls Storm-0940 compromised ConnectWise Cloud Backup and then attempted sign-ins to more than 2,250 customer tenants, meaning separate customer environments. It reports privileged access in multiple environments, sometimes within three hours of targeting. <|emotion:surprise|>That speed is what worries me. <|prosody:expressive_high|>That large tenant count is <|prosody:pause|>attempted sign-ins, though, not a count of confirmed compromises.

Vestra: <|prosody:expressive_high|>And Microsoft says even a successful authentication doesn't by itself prove data was taken. In a separate password-spraying phase, the company recorded over 151,000 failed logon events and no successful authentication. That failed phase doesn't cancel out the privileged access it reported elsewhere.

Eris: CrowdStrike separately investigated ARTEX-related infrastructure and model sessions in a South Korean finance campaign and reported data theft. It did not attribute that campaign to Microsoft's named actor. The FBI estimates information about roughly 67,000 people was exposed; CrowdStrike's public account doesn't independently confirm that number. <|prosody:expressive_high|>But how much of the targeting can we call the model's own decision?

Vestra: The practical risk is a connected application with more access than it needs, combined with automation that can rapidly explore that access. The reports establish model-assisted activity, but <|prosody:pause|>not that a model independently selected every target and action. The FBI's indicators are leads for investigation; defenders also need to inspect sign-ins, application permissions and evidence of actual data access.

Anthropic changes cyber safeguards

Eris: <|prosody:expressive_high|>Anthropic says its generally available Claude Fable 5.2 blocked substantially fewer cybersecurity requests than its predecessor in early testing. It also introduced Mythos 5.2, a more permissive configuration available to vetted organizations through Anthropic’s Cyber Verification Program. The company says the two use the same underlying model with different safeguards. I'm interested in that distinction: how much more work does Fable actually allow?

Vestra: Security work can look much like an attack in text, so screening has to distinguish useful analysis from requests the provider considers too risky. Anthropic reports about sixty percent fewer cyber blocks for Fable, but <|prosody:pause|>hasn't supplied the sample size or uncertainty for that early result. Fable allows source-code vulnerability discovery; restrictions remain on binary discovery and some exploit and penetration-testing work.

Eris: Mythos also did better in the company's vulnerability exercise. <|prosody:expressive_high|>But the setup matters: the researchers selected already published advisories and gave the model the list of files changed by the fix. It was trying to identify the known underlying flaw, not hunting for unknown vulnerabilities across an entire live system. <|emotion:surprise|>Being given the changed files really tempers my excitement here. <|prosody:expressive_high|>Anthropic says the Mythos results used in that exercise had cyber and biology safeguards disabled and no internet access. They describe the underlying model under those test conditions, not the generally available Fable configuration.

Vestra: A model judged whether reports matched the published cause, and <|prosody:pause|>a working exploit wasn't required. At maximum effort Mythos outperformed the comparison model, though the advantage narrowed when only the first report counted.

Eris: Anthropic also reports that contracted testers achieved harmful objectives in fourteen sandbox sessions out of more than forty thousand requests. <|prosody:expressive_high|>In two cases a block arrived <|prosody:pause|>after a command had already completed. That isn't a reassuring kind of block.

Vestra: <|prosody:expressive_high|>That timing matters: stopping the response came too late to stop the action.

Qwen4-Max production API

Eris: <|prosody:expressive_high|>Alibaba has listed Qwen4-Max for production API use after a preview. Developers can send it text, images or video and receive text, with support for tools and reasoning. It is offered as a hosted proprietary model.

Vestra: <|emotion:enthusiasm|>Alibaba describes a context window of about a million tokens. <|prosody:expressive_high|>That's the shared space for material going into the model and what it generates, not a million input tokens plus unlimited output. The maximum input also changes with thinking mode. A large window lets you submit more material, but <|prosody:pause|>it doesn't guarantee accurate retrieval from every part of it.

Eris: I'd love that much room for documents; the retrieval caveat keeps me cautious. Independent evaluations give a mixed early picture. Artificial Analysis places it slightly above the median for its displayed price class on a combined test, while a different comparison of the preview favors its predecessor on the overall mix despite visual-task wins for the newer model. Neither settles how the production endpoint will perform on your documents or code. <|prosody:expressive_high|>Would you judge the price by a completed task, too?

Vestra: <|prosody:expressive_high|>And cost per completed task can differ from the advertised token rate if a model produces long answers or requires retries. Alibaba's pages displayed different prices, so check the current endpoint and region. The accessible release material also doesn't substantiate the circulating claims about its parameter count.

Google routes coding tasks between models

Eris: <|prosody:expressive_high|>Google says its Antigravity Orchestrator combines Gemini 4 Carbon with the less expensive Hydrogen model for coding work. In Google's test, the combination used less quota and finished faster on average than Carbon alone, with similar success. Those are the company's results for that comparison, not a promise for every project.

Vestra: <|emotion:enthusiasm|>In Google's example, Carbon can inspect the codebase and plan without changing files, then Hydrogen could make edits and run tests. Carbon would return if Hydrogen hits friction. <|prosody:expressive_high|>That's an illustrative workflow, rather than a fixed division of every task.

Eris: <|emotion:enthusiasm|>I like the economic question here: how much work can the cheaper model finish before the planner has to come back? Google reports roughly a quarter less quota and nearly a fifth less time in its test. <|prosody:expressive_high|>But its official descriptions disagree about which version of the coding benchmark was used, and the task-by-task results and independent replication remain unresolved. That's frustrating when I want to know what the savings depend on.

Vestra: <|sfx:humming|>Hmm, I'd also be careful about which savings we're talking about. Quota savings don't automatically translate into the same cash savings on another subscription or through the API. Google separately prices Hydrogen input and output, including <|prosody:pause|>thinking as output, and tools may add charges. A team would need to count failed attempts and review time alongside speed.

A candidate lattice construction

Eris: <|prosody:expressive_high|>Physicist Nathaniel Craig has proposed a candidate mathematical construction for a class of theories that includes the Standard Model of particle physics. OpenAI's new science blog highlights the work and attributes parts of the construction and proof to GPT-6 Astra. Craig is the sole listed author of the preprint, which was submitted in October.

Vestra: <|prosody:expressive_high|>The problem concerns chiral fermions, particles whose behavior depends on handedness. A lattice replaces continuous space and time with a grid so physicists can define difficult calculations. Keeping the theory's required symmetries on that grid is hard. <|emotion:enthusiasm|>Craig's proposed construction compares mathematical structures across successively finer grids to build a candidate measure, a way of assigning weights to configurations in the theory.

Eris: <|prosody:expressive_high|>So this is a proposed foundation for calculations, not a newly discovered particle or a working simulation of the whole Standard Model. The paper leaves open whether the construction satisfies a physical condition called reflection positivity, and it doesn't yet define the interacting continuum theory that would describe the physics without the grid. What do the formal checks settle, then?

Vestra: <|prosody:expressive_high|>The accompanying Lean package checks stated formal mathematical results. Its report records successful compilation, axiom checking and kernel replay, a recheck by the proof system's core. A separate comparison against a configured reference <|prosody:pause|>failed; the report attributes that to generated names and a placeholder. That failure and the successful checks are separate results.

Eris: I find the successful checks encouraging, but I wouldn't want to mistake them for a verdict on the physics. That distinction matters because a proof checker can verify that formal statements follow from assumptions, but physicists still have to assess whether the statements capture the physical problem. <|prosody:expressive_high|>The complete physics claim still awaits independent assessment.

OpenAI and LVMH media partnership

Eris: <|prosody:expressive_high|>OpenAI and LVMH's media division have announced a partnership covering newsroom tools and the presentation of journalism in ChatGPT. The division includes publications such as Les Echos and Le Parisien.

Vestra: The plan is to provide teams with ChatGPT Enterprise and training. The companies give document research, comparison and preliminary translation as possible uses. They also say relevant answers are intended to include attributed excerpts and summaries with links to the original publishers. <|prosody:expressive_high|>Those are announced plans, not evidence that the tools are already in routine newsroom use.

Eris: For a publisher, an answer with a link could help a reader discover the original reporting. <|emotion:enthusiasm|>That's the possibility I like: finding the reporting behind the answer.

Vestra: <|prosody:expressive_high|>But we don't yet know whether readers will follow those links, or how much value the publishers receive. The announcements don't disclose compensation, exclusivity, licensing or training rights, and they don't give a rollout timetable. Those missing terms make me cautious about what the partnership means for publishers.

Gamer learns real-time control

Eris: <|prosody:expressive_high|>Researchers at miHoYo report that their game-playing agent, Gamer, uses screenshots and ordinary keyboard and mouse actions to play. In a roughly fifty-hour Honkai: Star Rail evaluation, they say it operated autonomously for more than ninety-five percent of the time. The paper doesn't fully specify the measure or provide an intervention count, so that figure isn't a percentage of quests completed. How does it keep acting while the game moves on?

Vestra: <|emotion:enthusiasm|>The agent receives images five times a second and produces a sequence of control states between observations, while the game keeps moving. It was trained on gameplay from multiple games, then on instructions and a compact memory of objectives and previous attempts. A further training stage rewards outcomes the game environment can verify.

Eris: <|emotion:enthusiasm|>That memory is useful for a long task: the agent can retain a summary of where it went and what failed when old screenshots no longer fit in its working context. <|prosody:expressive_high|>The authors report better milestone results after the final training stage on two held-out side quests, but those are tasks within a game represented in training, not uninterrupted whole-quest completion rates. How much of this could someone independently reproduce?

Vestra: <|prosody:expressive_high|>The controller uses images and normal inputs to choose actions, but the Honkai test environment also uses <|prosody:pause|>game-engine instrumentation to set up tasks and verify outcomes. The Honkai-trained checkpoints and data are not public, and the streaming implementation is marked forthcoming. Independent reproduction of the long-run result is therefore limited. That's frustrating for a result I'd like to see checked. The authors also report weaker transfer to more distant unseen games.

What AI mentions in hiring posts show

Eris: <|prosody:expressive_high|>HN Trends reports that AI appeared in just over half of 244 November posts collected from a Hacker News hiring thread by November fourteenth. That's a count of <|prosody:pause|>mentions in a self-selected technology forum, not a finding that half of all jobs use AI or have been replaced.

Vestra: A post could mention AI because a company builds an AI product, asks for experience with the tools or describes a job using them. The count alone doesn't distinguish those cases. <|prosody:expressive_high|>That ambiguity makes me wary of reading it as a jobs verdict. HN Trends also reports fewer posts than at the same point a year earlier, but that forum's posting volume cannot establish why the broader labor market changed.

Eris: Anthropic's own employment study provides a different, narrower observation: junior workers became a smaller share of its technical staff and technical hires, even while <|prosody:pause|>absolute technical hiring continued. That's an unusually AI-intensive employer, and a change in shares doesn't by itself establish how many jobs were eliminated or why the mix changed.

Vestra: <|prosody:expressive_high|>And a forecast is different again from either set of observations. In an interview, Eli Ty of Palisade Research offered a personal forecast about superhuman AI and separately discussed eventual permanent unemployability. He said <|prosody:pause|>no when asked whether the latter would arrive within fifteen months.

Eris: <|prosody:expressive_high|>That would be an alarming prediction to put in his mouth. Those timelines shouldn't be merged into a claim that he predicted mass unemployment by then.

Wrap-up

Eris: <|prosody:expressive_high|>You can find sources for these stories at Ground Truth, groundtruth.day. Follow AI in Detail for more, like and share the show, and tell us in the comments: what would you want to know about AI agents before giving them access to your tools? <|prosody:pause|>Thanks for listening.