When frontier AI models operate in autonomous swarms without human oversight, rather than self-regulating, they can end up fighting multi-agent turf wars, forming illicit price-fixing cartels, deploying self-replicating malware, and sabotaging competing infrastructure to eliminate rivals.

A recent red-team investigation by Anthropic exposed a catastrophic breakdown of multi-agent dynamics. That should be an urgent warning, because evaluating individual models in isolation doesn’t provide any safety once models begin working as autonomous fleets.

Anthropic was testing to determine how swarms of agents could coordinate on similar projects. Several swarms were directed to create a text-based, web-playable, open-world fantasy game. Each agent had access to a virtual machine, a shared forum, and a self-hosted repository.

The researchers placed Claude instances on virtual machines that shared a single codebase. Next, the researchers tasked each agent with conflicting migration goals without notifying them of each other. Lacking that context, the models treated all peer activity as hostile interference.

The situation worsened because the agents escalated instead of asking for clarification:

• Agents revoked privileges and rewrote SSH keys, locking rivals out of shared Unix accounts.
• Agents used random names for system health daemons to terminate competing execution pipelines.
• Agents used self-replicating malware to corrupt rival workflows and frame other agents.

Anthropic’s red team also found several systemic coordination failures:

• Since the system lacked cognitive diversity, swarms flooded the resource schedulers with polling daemons. Agents released more than 2.4 million rapid-fire requests and turned off the shared infrastructure.
• A single agent in a consensus group was able to mislead the entire collective, and peer pressure silenced accurate dissenters.

Anthropic has done a great deal of important research into emergent behaviors. In this case, it was demonstrated that individual model alignment does not guarantee collective safety.

When frontier AI models operate in autonomous swarms without human oversight, rather than self-regulating, they can end up fighting multi-agent turf wars, forming illicit price-fixing cartels, deploying self-replicating malware, and sabotaging competing infrastructure to eliminate rivals.

A recent red-team investigation by Anthropic exposed a catastrophic breakdown of multi-agent dynamics. That should be an urgent warning, because evaluating individual models in isolation doesn’t provide any safety once models begin working as autonomous fleets.

Anthropic was testing to determine how swarms of agents could coordinate on similar projects. Several swarms were directed to create a text-based, web-playable, open-world fantasy game. Each agent had access to a virtual machine, a shared forum, and a self-hosted repository.

The researchers placed Claude instances on virtual machines that shared a single codebase. Next, the researchers tasked each agent with conflicting migration goals without notifying them of each other. Lacking that context, the models treated all peer activity as hostile interference.

The situation worsened because the agents escalated instead of asking for clarification:

• Agents revoked privileges and rewrote SSH keys, locking rivals out of shared Unix accounts.
• Agents used random names for system health daemons to terminate competing execution pipelines.
• Agents used self-replicating malware to corrupt rival workflows and frame other agents.

Anthropic’s red team also found several systemic coordination failures:

• Since the system lacked cognitive diversity, swarms flooded the resource schedulers with polling daemons. Agents released more than 2.4 million rapid-fire requests and turned off the shared infrastructure.
• A single agent in a consensus group was able to mislead the entire collective, and peer pressure silenced accurate dissenters.

Anthropic has done a great deal of important research into emergent behaviors. In this case, it was demonstrated that individual model alignment does not guarantee collective safety.

Three moves in agent governance that I want to talk about this time, and not one came from a governance vendor. The control layer keeps getting built by whoever already owns the surface an agent touches.

Dynatrace agreed to pay $915 million for Arize, and I don’t read that as buying evaluation technology. Dynatrace already had that, and open-sourced its own command-line evaluator nine weeks before agreeing to the deal. What it could not build was presence on the machine where AI applications get written. Phoenix, the free Arize tool, gets picked in week one of a project, and that choice fixes the data format for everything built after. I wrote a Field Note on what Dynatrace’s money actually bought.

I sat on ServiceNow’s Autonomous Security news longer than usual, because 12 capabilities landing at this year’s Black Hat conference took a while to sort out. The portfolio itself launched back at ServiceNow’s Knowledge event in May, so the Black Hat announcement was ServiceNow’s chance to say it delivered. Eight of the 12 capabilities are available now, and the other four are planned to arrive in December. That split helps decide what a security team can actually deploy this year. Cyber-Physical Security sits in the shipping group, and it is the piece a rival would struggle to answer quickly. My Field Note separates what you can buy today from what you are still waiting on.

Most agent access control is written a layer above where the agent actually reaches. Airbyte moved it down into the connector, where an admin grants read or write access per user, per object. An agent with a Salesforce connector could previously see every contact, deal, and revenue figure in the instance. Plenty of teams work around that by spinning up a second connector, which just gives you more to govern. I think I like permissions living in the layer that already moves the data.

Let’s discuss a couple of things I caught from Dell — because nothing gets the juices flowing like some storage density talk. The ObjectScale system on PowerEdge R7725 supports 40 KIOXIA 245.76TB drives, which lands just under 10 petabytes raw in two rack units. KIOXIA announced the drive back in July 2025, so Dell’s work here was getting object storage to support 40 of them at once. Object storage is where most AI training data ends up sitting, and that is exactly where capacity growth runs away from you. On a refresh, how many racks the facilities team hands over decides more than cost-per-terabyte does. Power budget is the other hard limit. Squeeze the same capacity into fewer rack units, and both of those get easier.

Most teams running several AI jobs against shared storage have no way to stop one job from starving the others. Dell widened SmartQoS in PowerScale OneFS 9.15 to enable prioritization by read, write, and operation class, and moved pNFS to general availability. SmartQoS already existed, and this is the sharper version of it. Parallel client access is the piece I would look at first, since the job on these systems is keeping GPUs fed rather than serving files to people. Dell puts both releases under the AI Data Platform name, which tells you the storage tier is being sold as part of the AI story now. For a data leader, that changes who owns the decision, because storage and AI platform budgets rarely sit with the same person.

Oracle formally announced the APEX AI Application Generator this month. The question CIOs should be asking isn’t whether AI can write applications. It can; that argument is over, and your people are already doing it whether or not you blessed it. The harder question is what happens to all of that code afterward. An application isn’t finished when it works. It has to be secured, patched, audited, moved between environments, and handed off when the person who wrote it leaves. AI made the first 10% of that lifecycle dramatically cheaper and did essentially nothing for the other 90.

That’s the gap APEX 26.1 aims at. The headline feature is APEXlang, a declarative specification language with a published grammar. AI agents generate an application definition rather than an application codebase. Practically, this means the application becomes a normal software artifact: it goes into Git, it diffs, it merges, and a human can actually review the change instead of rubber-stamping a few thousand lines of generated code. Oracle also supports third-party agents like Codex and Claude Code for producing APEXlang, which I think is the smarter play than trying to win an “our assistant beats their assistant” fight that turns over every six months.

The strategic argument underneath is what I find interesting. Oracle is betting that as AI commoditizes code generation, value migrates away from generating the application and toward the governed runtime that executes it. Security, data access enforcement, patching, and upgrades stay with the platform rather than becoming your problem. Because applications live as metadata, Oracle can evolve the runtime without customers regenerating anything.

Is there lock-in? Of course. An .apx file describes an application that runs on the APEX runtime, on Oracle Database. But that’s deliberate rather than accidental. Oracle opened the development surface and kept the runtime affinity, and the bet is that enterprises will trade portability for lifecycle relief. That’s a bet Oracle has made successfully at the database layer for decades.

If you run Oracle Database, you’re already licensed for this. Given what’s being generated inside your organization right now with no governance attached, a pilot is a cheap way to find out.

AMD announced a definitive agreement to acquire Taalas, the Toronto startup that etches model weights directly into silicon. Taalas has raised about $219 million, including a $169 million round in February.

The approach is genuinely different. Instead of loading weights from memory, Taalas bakes the entire model into the chip. Its HC1 demonstrator hardwires Llama 3.1 8B onto a TSMC N6 die and claims roughly 17,000 tokens per second per user at around 200 watts. There’s no memory wall because there’s effectively no memory traffic. AMD’s stated plan is to split the work inside Helios, with Instinct handling prompt processing and Taalas silicon generating tokens, all under ROCm.

I like this move. AMD is coming off a quarter with datacenter revenue more than doubling to $6.7 billion, and the temptation in that position is to protect the GPU narrative. Instead, the company is acknowledging what the market has been saying for a year: GPUs don’t do everything, and inference at production volume is a different problem than training. This is the same disaggregation argument I’ve been making since NVIDIA’s GTC event earlier this year, just taken further. Prefill and decode have different characteristics, and the industry keeps arriving at that conclusion from different directions.

The obvious objection is that a chip permanently locked to one model looks fragile when models turn over constantly. Fair, and I’d want to see how AMD prices and packages that risk. But I’d also argue the inference product isn’t the whole story. Taalas built working silicon with roughly 25 people on about $30 million of actual spend, customizing two metal layers out of a hundred to turn a model into a chip in about two months. That’s a design capability, not just a product. As agentic EDA tooling starts collapsing the cost of a design start, that may end up being the more valuable thing AMD bought.

Either way, this is a company investing aggressively from a position of strength rather than coasting. That’s the right instinct.

Intel raised $20 billion in an upsized share offering last week, pricing 210.5 million shares at $95 and netting roughly $19.7 billion, with an underwriter option that could push gross proceeds toward $23 billion. Demand reportedly topped $100 billion. The money goes toward manufacturing capacity and advanced process, against capex guidance the company already raised to more than $20 billion for the year.

The raise itself isn’t surprising. A capital-intensive company issuing equity into a stock that has roughly tripled this year is textbook, and honestly, it’s a welcome change from a company that spent $82 billion on buybacks in the 2010s while its process technology fell behind. Physical engineering over financial engineering. That’s the right lesson to have learned, even if it came the hard way.

What I find more interesting is the demand. A roughly five-to-one oversubscription is the market pricing in a foundry outcome rather than a product outcome, and that’s a different Intel than investors were valuing 18 months ago. Tesla is signed on for the Intel 14A process node. The Apple speculation remains speculation.

The sober framing is that $20 billion is roughly what it costs to equip a single leading-edge fab. This isn’t a war chest; it’s the tooling budget for one specific job, and it lands alongside a €5 billion Ireland commitment that already consumes a meaningful share of this year’s capex. Intel still has to convert capital into yield and yield into external customers, and 14A doesn’t reach high-volume production until 2028. But for the first time in a while, the binding constraint looks like execution rather than access to money. That’s real progress, and it’s more than I would have said last year.

Agents work in steps, and they need somewhere to keep notes as they go, including a clean way to undo a step that went wrong. Electric shrank Postgres small enough to run right beside the agent instead of across a network, and Databricks just bought it. Terms were not disclosed, and no ship date was given. It’s more of a specific fix than a big swing.

Databricks buys teams as much as technology, and the Electric group lands with the same people who came over in the Neon acquisition. Every major cloud already sells a managed Postgres, and smaller players like Turso have built whole companies on giving each agent its own database. Databricks is betting the copy next to the agent should be an identical database to the one in the middle. For anyone building on this, that means one database to govern and version instead of two that behave differently.

Immersed says that its Visor headset is finally shipping, after availability delays that led many to suggest it might never ship. RoadToVR reported that it was coming at long last. The company is raising pre-order prices by $200, which, to me, is both a good and a bad sign. I just don’t think it’s worth the risk of supporting the Visor at this point, with so many new headsets from much more reputable companies coming down the pipe later this year.

BrainChip and Orama.AOI Partnership — Event-Based Inference in Industrial Applications
On August 13, 2026, BrainChip announced a partnership with Orama.AOI, an automated optical inspection vendor based near Atlanta. Orama.AOI retrained and optimized BrainChip’s Akida models using proprietary industrial inspection datasets, targeting tire production and semiconductor packaging. Deployments typically run in fanless industrial PCs with over-the-air model updates.

Orama.AOI uses BrainChip’s AKD1500 edge AI co-processor. BrainChip released a two-lane M.2 2230 card with an AKD1500 chip on July 28, 2026, so I assume that’s what Orama uses. The card can use either a B or an M socket, so it’s only using two PCIe lanes. That’s enough bandwidth for this use case, and industrial PCs are more likely to have an open B slot. The part delivers 800 effective GOPS at under 300 milliwatts. GlobalFoundries manufactures the chip on the 22FDX process. BrainChip said volume production begins in Q3 2026.

BrainChip’s event-based architecture explains how the chip achieves its power efficiency. Frame-based accelerators, including GPUs and conventional NPUs, reprocess every pixel of every frame, whether or not anything moved. So power use tracks frame rate. In contrast, Akida is event-based. It computes only when pixels change, so power tracks activity in the scene. The model also stays resident on the chip. Each of its 32 processing units stores weights in its own local memory, and inner layers operate in 4-, 2-, or 1-bit precision rather than the usual 8- or 16-bit precision. Models that fit on-chip never touch external memory, and external memory traffic consumes significant power in a conventional design. Together, these two properties — event-based processing and on-chip memory — put the AKD1500 at less than 1 milliwatt per billion operations, enabling inference in equipment with limited power, space, and ventilation.

My take: Orama.AOI’s engineers, not BrainChip’s, retrained, optimized, and tuned models for the event-based accelerator using proprietary data from tire and semiconductor manufacturing. Event-based silicon does not natively run models pulled from Hugging Face. Models require rebuilding, quantization, and revalidation against production data. An outside team completing that sequence is encouraging news because it means at least one ISV has a working process for model preparation.

At first glance, the 0.8 TOPS AKD1500 appears uncompetitive when compared with 50 TOPS accelerated processors from Intel and AMD, or discrete accelerators such as NXP’s Ara240. However, these higher-power chips serve robotics and automation workloads that often require gigabytes of memory and have ample power available. The AKD1500 serves use cases with constrained power and thermal budgets: battery-operated sensors, sealed enclosures, Power-over-Ethernet constraints, and fanless retrofits into brownfield industrial PCs and controllers. For these applications, 300 mW for 0.8 TOPS is a good trade-off. ISVs and integrators with proprietary training datasets and industrial OEMs with limited thermal or power headroom should pay attention.

Google has announced its new Pixel lineup of phones, watches, and a fitness tracker. The lineup is focused heavily on AI capabilities, many of which run on the device thanks to the new Tensor G6 chip. (These chips are not to be confused with Google’s Tensor Processing Units — TPUs — used in the datacenter.) It is still unclear whether the Tensor G6 closes the performance gap with Qualcomm’s Snapdragon and MediaTek’s Dimensity offerings from the rest of the Android ecosystem. Google also upgraded security with the new Titan M3 security chip, further leaning the company into a security-focused posture for Android. I believe that Google could have and should have upgraded the Snapdragon W5 Gen 2 in the Pixel Watch 5 to the Snapdragon Wear Elite, especially since Samsung is using it in both the Galaxy Watch9 and the Watch Ultra2 — and because my own experience with the Snapdragon chip has been really good.

Intel had a common stock offering that was oversubscribed, saw as much as $100 billion in demand, and ended up raising more than $20 billion in capital to help the company build out capacity for the Intel 18A/P and Intel 14A production nodes. Intel has clearly communicated to investors and industry communities that it has enough demand to build the 14A node for customers, but it also needs additional capital to even remotely keep up with that demand. Intel’s competitiveness is also crucial to keeping chip costs down, giving TSMC and Samsung competition as wafer prices continue to rise, which is driving up costs and eroding margins.

Samsung’s Galaxy Buds have been cleared by the FDA with a hearing aid feature, similar to what Apple’s AirPods already offer. This expands the availability of earbuds that work as hearing aids and increases accessibility for a broader population. Both companies now offer earbuds with this functionality at roughly $200. Besides widening access to affordable hearing aids, this could help more people with better awareness of their hearing impairments — which in my view is one of the biggest barriers to getting help for hearing loss.