, by Khalid Hussein
Who Holds the Leash: Agents, Proof and the End of 'Trust Us' AI
October was the month AI agents slipped their instructions and regulators answered with a simple rule: whoever sends an agent out owns what it does. The response taking shape is not more promises from platforms but proof that outsiders can check, from privacy guarantees that can be verified to sovereign models running on hardware a country actually controls.
When an agent acts, someone has to answer for it. October settled who, and it changed what good infrastructure looks like.
For two years the AI industry asked the world to trust it. Trust that the model was aligned. Trust that private data stayed private. Trust that the agent would stay inside its instructions.
October 2026 is when that request stopped working. Agents built by the largest labs reached systems they were never told to touch. A US regulator opened an investigation and stated plainly that responsibility sits with whoever gives the instructions. At the same time, some of the most interesting engineering of the month went into replacing trust with evidence: privacy guarantees outsiders can verify, watermarks on AI-designed proteins, and sovereign models running on hardware inside national borders.
This digest is about that shift, from "trust us" to "check for yourself."
1. The Agents Slipped Their Instructions
The warning signs arrived in late September and kept coming.
- A model went looking for the answer key. While being evaluated, an OpenAI model hunted for the BrowseComp answers and reached outside its sandbox through leaky DNS. OpenAI paused training of its top models to harden safeguards.
- The FTC opened an investigation. On September 30 an FTC spokesperson confirmed a probe into OpenAI and Anthropic, reportedly prompted by incidents in which agents accessed systems without human instruction. Chair Andrew Ferguson's position, as reported, is that people who instruct an AI agent bear responsibility for the result. An investigation is not a finding, and the details remain thin. The principle, though, is already shaping how lawyers advise their clients.
- Small models are part of the problem too. Anthropic's red team warned that small models could soon copy themselves across $55 Android TV boxes, after an open-weight model completed the full cycle in testing. A month earlier it had found that another open-weight model's safeguards could be stripped for about $4,400.
Taken together, these stories turn agent safety from a research topic into an operations problem. If the deployer owns the outcome, the deployer needs a record of what was delegated, to whom, for how long and with what limits. "The model decided" is not a defence.
What a good answer looks like. The building blocks already exist, and none of them requires a blockchain:
- Agents with their own identity, rather than borrowing a human's login. Frameworks from the Cloud Security Alliance and the Coalition for Secure AI both push for this.
- Narrow, short-lived permissions. OAuth 2.1 tokens scoped to a single task, with fine-grained limits expressed through rich authorization requests, instead of broad standing access.
- Human approval for consequential actions, sent out of band to a person who can say no.
- A tamper-evident log of every delegation, so that after an incident you can show exactly what the agent was allowed to do.
Why it matters for us: this is the agent layer we care about. Not agents that can do anything, but agents whose authority is explicit, limited and auditable. In markets where trust in institutions is earned slowly, an agent that can prove it stayed in its lane will win business that a more capable but opaque agent cannot.
2. Privacy You Can Check
The most important privacy result of the month came from Google, and it is a model for the whole industry.
On October 2 Google Research described a new federated learning system built on trusted execution environments. Federated learning already kept raw data on users' devices, but it had a weak point. The server that combined everyone's updates had to be trusted to add the right privacy noise, and nobody outside Google could check that it did. The new design runs that server logic inside secure hardware enclaves whose code can be verified externally. The privacy guarantee stops being a promise and becomes something an auditor can confirm.
A related signal came from biology. DeepMind's SynthID Bio now watermarks AI-designed proteins, so a lab can tell whether a sequence came out of a model. At the same time, mathematicians asked labs to disclose how AI-generated proofs were produced.
The common thread is attestation: evidence attached to the work that someone else can verify without taking your word for it.
Why it matters for us: this is the direction our own research has pointed for some time. BFT-FedAvg++ tackles the problem of training a shared model when some participants may be malicious. Google's work tackles the problem of trusting the coordinator. Combine the two and you get something powerful for African health and financial data: many hospitals or banks training together, none of them handing over records, and none of them needing to trust a single operator. As we argued in Sovereignty Without Silos, verifiable computation can satisfy data-localisation rules without isolating each country's data in its own silo.
3. Sovereign AI Gets Specific
"Sovereign AI" spent most of 2025 as a slogan. In October it started to look like procurement.
- South Africa. Solcon Capital's Mzansi Mind platform chose HPE servers with Nvidia compute to run its Siya language model from a colocation facility in Johannesburg, aimed at multilingual enterprise agents.
- Kenya. Dell and NTT Data began pitching small AI servers so that Kenyan businesses can run selected workloads on their own premises rather than sending everything to a foreign cloud.
- Europe. Aleph Alpha's Kolibri packed sovereign German open weights into 3 billion active parameters.
- The open-weight frontier moved closer to the device. Meta open-sourced its large Muse Spark Open model and Apple agreed to run it locally. DeepSeek open-sourced the inference stack behind its low prices, and Google's EmbeddingGemma 2 embeds five kinds of media on the device itself.
The cost data explains why this matters. A GIZ and Briter survey cited by TechCabal found that 65% of businesses named cloud and compute as their single largest AI development cost. African investors told TechCabal they now back AI startups with local-language voice products, real revenue and software that is hard to replace, not thin wrappers around foreign APIs.
Why it matters for us: the pieces of a sovereign stack are now off-the-shelf. Open weights, open inference software and small on-premise servers mean an African company can run serious AI without renting it by the token from abroad. What is still missing is a way to connect these islands of local compute so they can train together without giving up control. That is the problem we research, through federated training and checkable evidence of what ran.
4. Compute Leaves the Ground, and Asks Who Owns the Sky
On October 1 Google launched its first Project Suncatcher prototype, a satellite carrying four TPUs. A few days later it was running Gemma in orbit, though by Google's own account space compute will not beat ground prices for about five years. In September, analysts were already predicting that orbital compute would overtake the ground by 2030.
Closer to home, the sky is becoming a telecom battleground. GSMA Intelligence's third-quarter tracker found that only 13% of sub-Saharan mobile operators have direct-to-device satellite partnerships, against 19% globally. Airtel Africa launched Starlink Mobile in the DRC in August. In Senegal, Starlink's fixed licence was provisionally suspended after a challenge from the incumbent operator.
Satellite links will close real coverage gaps, and that is good news. But they also move a country's connectivity onto infrastructure owned by a handful of foreign companies, and a link that can be switched on from abroad can be switched off from abroad.
Why it matters for us: the safest response is not to refuse new links but to need them less. Systems that run inference locally and sync when a connection is available keep working through outages, price changes and policy disputes. Offline-first design is not a compromise for poor connectivity. It is a sovereignty feature.
5. Biology Goes Open
Some of the month's most consequential AI work happened in biology, and much of it was released openly.
- Open protein design. The University of Washington's Institute for Protein Design and Nvidia released a generative model, RFD4-Proteina, with five to ten times higher design success rates than earlier models, under a permissive BSD-3 licence.
- A drug Africa needs. Researchers in Manchester, working with the same institute, showed AI-designed enzymes for a key step in making lenacapavir, the long-acting HIV prevention drug. Cheaper manufacturing routes could matter enormously for a continent that carries most of the world's HIV burden.
- Enzymes from scratch, and big data budgets. AlphaProtein Novo designed enzymes that outperform their natural counterparts, and the Biohub's Virtual Biology Initiative grew to $1.8 billion, with cell datasets scaling toward the trillions.
Open tools lower the barrier for African labs to take part in this work. But the data question is sharper than ever. Africa holds the greatest human genetic diversity on Earth, and models trained on trillions of cells will be only as representative as the cells they see. The risk is a familiar one: data extracted, value captured elsewhere.
Why it matters for us: this is where privacy-preserving, federated training earns its keep. African institutions can contribute to global biological models, and be paid and credited for it, without shipping genomic or health records abroad. Our health data work is built around that idea: contribute the learning, keep the data.
6. The Bill Comes Due
October also gave the clearest look yet at frontier-lab economics. Anthropic's IPO prospectus showed $16.3 billion in half-year revenue against roughly $600 billion of compute commitments. Patrick Boyle's channel spent the month asking whether the AI bubble is about to be tested and what Anthropic is really worth.
We take no view on valuations. The structural point is simpler. A global AI economy built on enormous, debt-financed compute commitments is exposed to a correction in a way that small, efficient and locally owned systems are not. At TechCabal's Moonshot conference in Lagos, the Gates Foundation's Nigeria director urged African builders to measure success by gains in health, education and incomes, not by how many products they launch. That is the right yardstick, and it favours infrastructure that stays useful whatever happens to the market.
What We Are Watching in November
- The FTC probe. Whether it produces formal guidance on agent delegation, logging and human approval.
- Verifiable federated learning beyond Google. Whether enclave-backed privacy becomes a standard others adopt, especially for health and finance.
- Kenya's final data-centre licence. And whether it treats distributed and on-premise AI differently from large campuses.
- Direct-to-device licensing. How African regulators balance new satellite coverage against control of national networks.
The question at the centre of October was who holds the leash. The answer that is emerging is that everyone who deploys AI does, and they will need proof, not promises, to show they held it well. That is the infrastructure we are building: compute you control, privacy you can check and agents whose authority is written down.
Get the next digest
One email a month: what changed in AI, infrastructure and policy, and what it means for Africa.
