Agents Move From Chat to Labs, Sandboxes and Silicon

An AI-Found Enzyme Mystery

An AI agent sifted through more than 200,000 enzymes and surfaced something human scientists could verify in the lab. Then ten attempts to repeat the search all missed it.

Our main story today is Anthropic's 23 September claim that Claude agents found an uncharacterised enzyme system, and what that result says about using agent swarms for scientific discovery. Anthropic has launched an in-house life-sciences research group and laboratory. For its first reported project, it gave Claude agents a high-level research brief and asked them to search bacteriophage sequence data. Bacteriophages are viruses that infect bacteria, and their genetic data contains an enormous number of patterns that may be biologically interesting but haven't been characterised.

Anthropic says roughly 950 agent sessions searched more than 200,000 reverse transcriptases over about 21 hours, consuming approximately 210 million tokens. A reverse transcriptase is an enzyme that copies RNA into DNA. The agents identified a group of array-associated reverse transcriptases, or ARTs, sitting beside unusual repeated sequences. Human scientists then took over and confirmed that the associated repeat array is expressed as distinct short RNAs. That experimental result makes the computational lead biologically real enough to investigate. It doesn't reveal what the system does.

The handoff between machine search and laboratory work is the strongest part of the result. The agents reduced a huge sequence-mining problem to a specific system scientists could test. The scientists then established one observable fact about its RNA. Each stage answered a different question, and neither could supply the missing biological function.

That distinction is crucial. The enzyme's biochemical activity and biological function are still unknown. Anthropic describes the repeats as CRISPR-like, but the work hasn't demonstrated a gene-editing tool, a defence system or a practical application. The technical report hasn't been peer reviewed, and no independent group has replicated the finding. At this stage, the discovery is a verified pattern with an unresolved function.

The awkward detail is reproducibility. The report says ten repetitions of the campaign all failed to rediscover the defining array. That doesn't erase the first result; the laboratory evidence still exists. It does mean the search process can't yet be treated like a dependable scientific instrument. A conventional database pipeline may miss an odd connection because it follows fixed rules. A swarm of agents can explore more loosely, compare clues and pursue anomalies, but that flexibility also makes the route to a result harder to reproduce.

That split gives research groups a better way to judge the result. The laboratory observation and the agent campaign pose different reproducibility questions. Scientists can test again whether the repeat array produces those short RNAs. Separately, engineers need to ask whether the same research brief and search process reliably surface the same candidate. The first has supporting experimental evidence; the second failed ten times in the report. An agent workflow may still be valuable as a generator of unusual leads, but its hit rate can't be inferred from one success.

For research organisations, the sensible interpretation is benefit with a firm limit. Agents can widen a search across data at a scale that would be tedious for people and may find leads that fixed pipelines overlook. The 210-million-token run also shows that this kind of exploration has a substantial compute footprint, even before lab work begins. Expert review decides whether the reasoning holds, and experiments decide whether the biology is real. Until another run or another research group can reliably find the same system, ART is a promising lead, not scientific closure. The practical opportunity is a new way to generate hypotheses; the operational challenge is preserving enough evidence, repeatability and human scrutiny to know which hypotheses deserve scarce laboratory time.

An F5 Flaw Under Active Attack

There’s less room for patience in the next case: an access-gateway flaw is already being exploited.

F5 disclosed hotfixes on 22 September for CVE-2026-94127, a heap-based buffer overflow in BIG-IP Access Policy Manager. Under a specific configuration, an unauthenticated attacker can use it for remote-code execution. That means running code on the gateway without first signing in, which is about as serious as an internet-facing access-control problem gets. A heap overflow corrupts memory used by a running process. Here, the security consequence isn't a crash alone; the disclosed impact includes an attacker's code executing on the appliance.

The affected condition is narrower than every BIG-IP APM deployment. The device must be configured as an OAuth authorisation server, with an access policy and an OAuth profile attached to the same virtual server. Deployments used only as OAuth clients or resource servers aren't exposed through this particular attack condition. That distinction should help administrators identify the urgent population quickly, but it shouldn't soften the response for systems that do match. F5 told national cyber authorities that exploitation is happening in the wild. It hasn't publicly said how many systems have been targeted or who is behind the attacks.

The narrow trigger is operationally useful only if configuration records are accurate. An asset list showing that BIG-IP APM is installed doesn't reveal whether the appliance acts as the authorisation server, or whether the policy and profile share a virtual server. Administrators need that configuration-level answer before they can classify exposure. Once an affected gateway is identified, the prescribed log and core-file review gives the security team defined evidence to check for signs that remediation came after compromise.

Canadian and Japanese cyber authorities recommend applying F5's engineering hotfix or the vendor's interim iRule. They also point administrators towards evidence that could reveal an earlier intrusion: OAuth failures, audit logs and TMM core files. Those checks matter because a successful patch changes what can happen from that point forward; it can't prove the gateway was clean before the fix arrived.

My view for organisations running BIG-IP is to split this into two jobs. First, identify the exact OAuth-authorisation deployments and remediate them urgently. Second, investigate each exposed system as a possible incident, using the available logs and core files rather than closing the ticket when patching finishes. That makes the work harder than routine maintenance, but active exploitation changes the standard of evidence. A protected device and an uncompromised device are two different claims.

Microsoft's Middle East Infrastructure Plan

Zooming out from one gateway, Microsoft is making a much longer bet on regional cloud capacity.

On 24 September, Microsoft announced plans to spend more than 10 billion US dollars in capital and operating expenditure across Kuwait, Qatar, Saudi Arabia and the United Arab Emirates through 2030. The package combines cloud and AI capacity with connectivity, sovereign-cloud services, business continuity and cybersecurity programmes. More than 400 million US dollars is earmarked for subsea and terrestrial connectivity.

For governments and large organisations in those markets, a bigger regional Azure footprint could improve access to compute, data residency options and resilience. The sovereign-cloud element is especially relevant where sensitive workloads need tighter operational or jurisdictional controls than a standard public-cloud deployment offers. Business-continuity and cybersecurity programmes may also matter as more essential services depend on the same digital infrastructure. The connectivity allocation links those ambitions to a basic constraint: new cloud capacity is only useful when customers and regions can reach it reliably.

But these are forward commitments, not installed servers or available services. Microsoft didn't publish a country-by-country spending breakdown, a construction timetable or the capacity customers will eventually receive. The headline number also combines capital spending with operating expenditure, so it shouldn't be read as a single block of new data-centre construction. Delivery will depend on future building, partnerships and regional conditions.

The useful judgement for buyers is to treat the announcement as a strong signal for procurement planning, not as capacity they can consume today. The programme may give regional customers more capable and resilient infrastructure. It can also bind national AI and cloud programmes more closely to one hyperscaler's platform, commercial terms and technical roadmap. Organisations weighing those services will need to value the near-term convenience alongside the longer-term cost of moving workloads, identity systems and operational knowledge somewhere else. The investment may reduce an infrastructure gap while increasing strategic dependency. Both can be true at once.

A Security Guide for Connected Industry

A cloud region is only part of the picture when the connected systems also control buildings, water or freight.

NIST released the initial public draft of Special Publication 800-82 Revision 4 on 21 September. This is its updated security guide for operational technology: the computers, networks and controllers that monitor or change physical processes. The revision is organised around Cybersecurity Framework 2.0 and broadens its coverage beyond familiar factory systems. It now speaks directly to building automation, water and wastewater, food and agriculture, freight rail, maritime systems, industrial internet-of-things devices and the growing convergence between operational systems and cloud services.

That last point reflects how these environments are changing. A controller on a plant floor may still need strict availability and safety protections, while its monitoring data, identity controls or maintenance tools connect into enterprise and cloud platforms. The draft adds guidance on asset management, network monitoring, zero-trust architecture and aligning operational-technology risk with enterprise risk management. In plain terms, operators need to know which physical assets are connected, observe what those connections are doing, limit trust between systems and make the resulting risk visible to the people making organisation-wide decisions.

It isn't a new mandatory rule. Public comments remain open until 30 November 2026, and the final guide may change. Adoption outside regulated industries or procurement requirements will remain voluntary. That draft status is useful, though, because operators can test the guidance against real constraints before the wording settles. A conventional IT control can be a poor fit when restarting a device interrupts a production line or when network latency affects a safety process.

For organisations that run physical systems, this is a timely architecture check. Compare the draft with current asset inventories, monitoring and cloud connections, then submit evidence where safety or availability clashes with an ordinary IT assumption. Waiting for the final publication may produce a cleaner checklist, but it gives up the chance to shape guidance around how these systems actually behave. The harder security job is no longer confined to a sealed industrial network; it sits across the physical process, the corporate network and the cloud service joining them.

A Petabit Route Across the Atlantic

The physical network underneath all that cloud activity is getting an ambitious upgrade of its own.

On 21 September, Meta announced Petal, a planned 7,000-kilometre subsea cable between the United States and France. Its design target is one petabit per second, using two-core fibre. An independent report says the planned architecture has 24 fibre pairs. Meta says the multi-core approach would double the capacity of today's leading transoceanic cables without requiring a proportional increase in power or physical infrastructure. NEC and Sumitomo Electric are development partners, and Orange is supporting the cable's landing in France.

Multi-core fibre puts more than one light-carrying core inside a strand. That can increase the amount of data moved through a cable system without scaling every part of the physical package in step. For AI training, cloud services, video and the ordinary traffic moving between North America and Europe, more capacity per cable and per power envelope is attractive. It may also change how operators plan redundancy and future data-centre connectivity on both sides of the Atlantic.

None of that capacity is available yet. Petal isn't expected to enter service until 2029, and the one-petabit target is a company-supplied claim that hasn't been demonstrated in operation. Construction, commissioning, reliability and commercial access all remain unresolved. Subsea projects have long lead times, so the announcement is useful planning information, not an immediate improvement to anyone's connection.

There is also an ownership question. A successful multi-core system could lift the bandwidth available from a single route and reduce pressure to add power and physical infrastructure at the same rate. At the same time, another major cable owned by a large platform company puts more of the internet's physical backbone under the control of the businesses that already generate and carry enormous volumes of traffic. For cloud and content operators, the near-term move is scenario planning: account for a large prospective increase in transatlantic capacity, but don't build a 2027 dependency around infrastructure due in 2029.

Chip Design Meets Advanced Packaging

Down at chip level, the design tools are being pulled closer to the manufacturing process and the package around it.

Synopsys and TSMC announced certified electronic-design-automation flows for TSMC's A14 process on 23 September. The work covers digital and analog implementation through sign-off, along with AI-assisted workflows for digital, analog and multi-die designs. One of the new pieces is an agentic workflow for co-optimising chiplet floorplans, where multiple dies have to be arranged and connected as one system rather than designed in isolation.

That packaging layer has become central to high-performance AI hardware. Designers need to bring compute dies, high-bandwidth memory and fast links together while controlling power, heat, signal integrity and physical space. Synopsys also reported N2P tape-out milestones for interfaces including PCI Express 7.0 and HBM4. Tape-out is the point when a completed design is handed over for manufacturing. The companies also demonstrated UCIe-A links at 32 and 40 gigabits per second on an N3P test chip using a CoWoS-S interposer, the package layer that connects the dies.

The practical benefit is earlier access to a tool chain that spans advanced transistors, chiplets, memory interfaces and packaging. That may help AI and high-performance-computing chip designers find integration problems before tape-out, when changes become painfully expensive. An agentic floorplanning system may also explore more physical arrangements than a designer could assess manually, although the announcement doesn't supply an independent productivity benchmark. It gives no delivery date for customer chips built with these flows either.

For chip organisations, the opportunity is faster design convergence across parts of the system that increasingly have to be optimised together. The trade-off is tighter coupling to one foundry process and one EDA stack. If the automation performs well, that integration can save time; it can also make tools, design data and manufacturing choices harder to separate later. The certification is meaningful readiness work. Claims about shorter time to market still need evidence from commercial projects.

What Changes for You

The most immediate changes land at a developer's desk, with one new place to run agents and one new way to work with compact local models.

Docker has released Cloud Sandboxes for developers running coding agents. A task can begin in a local microVM, then its filesystem can move to Docker-managed cloud compute and keep running after the laptop disconnects. Each sandbox gets its own microVM, secrets and network policy, and Docker provides preconfigured kits for several coding agents. That makes a long build, test run or repository task less dependent on leaving a laptop awake, while keeping isolation and network controls around each agent.

There are practical limits. Cloud sessions last up to 24 hours and cost from 7 US cents to 1 dollar 12 per hour for one to 16 virtual CPUs. Model inference is billed separately by the developer's model provider. Access requires the sbx tool at version 0.45.1 or later, a pay-as-you-go plan, and a Docker Personal or Pro account. Local and cloud secrets and policies are separate, so moving the filesystem doesn't remove the need to configure the cloud side. For working AI builders, durable compute becomes easier; the bill and Docker control-plane dependency become part of the design. Docker's isolation and secret-proxy assurances also remain vendor claims without an independent assessment supplied at launch.

On Apple Silicon, Hugging Face has added an initial route for loading packed llama.cpp-style GGUF checkpoints through the normal Transformers APIs. It reuses ggml Metal kernels instead of expanding the whole model in memory. A developer can use the compact checkpoint from Python and PyTorch, or through an OpenAI-compatible local serving interface, which makes it easier to inspect, evaluate and modify a quantised model without first converting it to full-precision weights.

This path is early and narrow. It needs Apple Silicon, MPS, recent PyTorch builds and Transformers from the main development branch. Architecture support starts with Qwen3.5 and compatible Qwen3.8 checkpoints. Hugging Face still recommends llama.cpp when efficient local inference is the only goal, and its performance figures use slightly different benchmark paths. The gain is access to familiar research tooling, not universal GGUF support or a guaranteed speed win.