Get involved
For people at AI labs
Most people building AI want it to go well. We're not here to cast anyone as the villain. A few practical steps you can ship now would make a real difference.
In the early days of aviation, every aircraft maker built its own cockpit gauges, radio frequencies, and landing rules. Pilots switching planes misread dials in bad weather, and competing airlines felt pressured to fly into storms so they wouldn't lose a mail contract. Aviation became safe only when engineers, pilots, and regulators agreed on shared flight recorders, standard transponders, and weather minimums that applied to every carrier alike.
People working inside frontier AI labs face a similar bind today. Most researchers, safety engineers, and policy staff want the systems they build to remain honest, steerable, and genuinely helpful to the public. Yet when every lab invents its own private logs and races a quarterly release clock, individual goodwill gets squeezed by competitive pressure.
That is why shared, verifiable standards help conscientious lab teams as much as they help the public. When every ship paints the same load line on its hull, no captain has to gamble with safety just to stay in business.
Six practical steps a lab team can ship now
1. Put a licence plate and a receipt on every outward agent action. When your model calls an external tool, sends an email, or books a transaction, don't let it act as an unmarked ghost. Attach a signed envelope stating which user or organisation authorised the task, what narrow lease it was given, and a tamper-evident receipt the user can keep on their own device. Backing open workload-identity standards and attested capability leases makes this interoperable across labs.1,6
2. Separate untrusted reading from tool execution. An agent that browses the open web or reads incoming messages in the same privileged loop that controls bank transfers or email is an accident waiting to happen. Treat incoming text like mail arriving at a quarantine desk: parse and summarise it inside an unprivileged sandbox, and require explicit, user-visible capability checks before any external tool fires.4
3. Measure and bound long-horizon autonomy, not just single-turn chat. A model that is polite in a three-turn conversation can still drift badly on the twelfth step of a multi-hour software errand. Extend your frontier safety frameworks to measure autonomous task horizons and multi-agent interaction risks, and build automatic checkpoint pauses where the agent must renew its lease with a human principal.2
4. Train and reward honest uncertainty over confident bluffing. Today, many post-training benchmarks still reward models for guessing an answer rather than saying I don't know, and here is what we would need to check. Like a country doctor who orders a lab swab before prescribing, an assistant that states its uncertainty plainly and pauses before irreversible actions is far safer than one trained to sound polished.5
5. Protect independent red-teaming from pre-publication vetoes. Internal evaluations are essential, but every kitchen smells its own cooking differently than a health inspector does. Give independent academic and public-interest evaluators stable, pre-deployment access to frontier models and agent harnesses, with a clear legal safe harbour and the right to publish findings without a commercial veto.3
6. Support portable user memory and clean exits. Users should never feel trapped inside one assistant because years of personal notes and preferences are locked in a proprietary cloud format. Let people export, inspect, and delete their preference profiles on their own hardware anytime, like taking your own medical chart when you change clinics.
How your organisation is evaluated in Shared Horizon
In our Shared Horizon institutional explorer, we evaluate over three hundred organisations, including frontier labs, cloud providers, open-weight builders, and safety institutes. Each profile looks at five public-interest dimensions: human oversight, provenance receipts, decentralisation, independent safety evaluation, and democratic voice.
Those profiles are a working map of concrete next moves, and we update them. Perhaps your team has shipped a new provenance standard, opened an evaluation harness, or updated a responsible scaling policy that our dossier missed. Send us the primary link and we will update the record openly.
What we can offer your team
We have no commercial product to sell, no venture investors to please, and no stake in which lab wins a benchmark race. We can review draft specifications for agent receipts and capability leases, test multi-agent failure modes in our simulation harness, and convene neutral technical discussions across labs, standards bodies, and civil society.
Write to us at contact@collectiveintelligencenetwork.org. A person reads every message, and together we can build cockpit instruments that keep people safely at the controls.
1. J. Salowey, Y. Rosomakho, and H. Tschofenig, Workload Identity in a Multi System Environment (WIMSE) Architecture, IETF Internet-Draft (2026); M. Jones et al., WIMSE Workload Credentials, IETF Internet-Draft (2024–2026); and A. Schwenkschuster and Y. Rosomakho, Workload Identity Practices, IETF Internet-Draft (2026).
2. Thomas Kwa et al. (METR), Measuring AI Ability to Complete Long Tasks, arXiv:2503.14499 (March 2025).
3. UK AI Safety Institute, Frontier AI Task and Deployment Evaluations (2024–2026); and Apollo Research, Auditing Agentic Deception and Scheming (2024–2026).
4. Edoardo Debenedetti et al., CaMeL: Defeating Prompt Injection via Capability-Based Access Control, arXiv:2503.18813 (March 2025).
5. Dylan Hadfield-Menell, Stuart J. Russell, Pieter Abbeel, and Anca Dragan, Cooperative Inverse Reinforcement Learning, NeurIPS (2016), arXiv:1606.03137.
6. Zhiyang Ding, Yang Luo, Guangpu Chen, Qingni Shen, and Zhonghai Wu, ACLE-MCP: Attested Capability Leases for Execution-Time Trust in Remote LLM Tool Use, arXiv:2609.02690 (September 2026).