Cloud Labs and Automated Biology
Some cloud-laboratory workflows let a team design experiments remotely, submit them to robotic facilities, and receive results without anyone from that team entering the lab. That model weakens four assumptions built into traditional biosecurity: trained personnel are physically present, institutional review covers the work, physical access controls limit who can run experiments, and colleagues observe what happens at the bench. Proportionate governance therefore centers customer verification, activity logging, prohibited-order lists aligned with DNA Synthesis Screening, and strict limits on AI systems that can call lab APIs - not a ban on legitimate remote research.
Cloud labs are a potential control-path loophole when KYC and API authorization are weaker than in on-prem BSL facilities.
Decision-makers who need the handbook-wide overview should start with the Executive Summary, which places cloud laboratory governance alongside DNA synthesis screening, AI evaluation, and BWC implementation priorities.
- Explain what cloud laboratories are and how they enable remote experiment execution.
- Identify the biosecurity assumptions that cloud labs disrupt.
- Analyze the convergence risk when AI systems gain access to automated lab platforms.
- Evaluate current governance approaches and their limitations.
- Apply a risk-proportionate framework for balancing research access with security.
Introduction
Imagine designing a biological experiment on your laptop in one country while robots in another country execute it automatically, with results streamed back in real-time.
This is not science fiction. This is cloud laboratories - and they are already operational.
Earlier chapters in this Part traced how AI affects biosecurity across multiple dimensions:
- AI as a Biosecurity Risk Amplifier introduced the framework: LLMs as information tutors, BDTs as capability enablers
- LLMs and Information Hazards examined how LLMs can provide troubleshooting guidance and aggregate dual-use knowledge
- AI-Enabled Pathogen Design analyzed how AI tools can design novel biological sequences
- AI for Biosecurity Defense showed how AI can strengthen detection and countermeasure development
The convergence of these capabilities with automated laboratory execution is a qualitative shift. Cloud labs could bridge the final gap: the “tacit knowledge” barrier that has historically limited who could actually conduct biological experiments.
Throughout this handbook, we have noted that the tacit knowledge barrier - the hands-on experimental skills required to successfully conduct biological work - remains one of the most important practical barriers to misuse.
You can read every protocol in the world, but making experiments work requires practice, troubleshooting, and physical technique. This is why the biosecurity community has, until now, focused primarily on information and materials rather than laboratory access.
Cloud labs change this calculus. If experiments can be designed remotely and executed automatically, the tacit knowledge requirement may begin to dissolve.
Headlines that treat every automated or remote wet lab as “autonomous AI biology” blur three axes that are not the same. Automated means the machine executes a pre-defined experiment. Autonomous means the system also makes decisions about the experiment (variable choice, next-step selection). Cloud means remote access to the physical laboratory. These are not mutually exclusive: cloud labs are often highly automated, and autonomous loops can run through cloud infrastructure. Steph Batalis’s What Does “Autonomous” Actually Mean in a Wet Lab? (30 Sep 2026) puts that vocabulary next to the Anthropic wet-lab coverage and argues the demonstrated systems still need narrow preselected inputs and heavy human oversight. Substack teaching essay; not a methods primary. Keep the Ginkgo preprint and Anthropic institutional posts as evidence cites.
On biosafety, the useful fences are operational rather than alarmist. Biology is not cyber: a compromised planner cannot create a real-world infection by sending a command alone; physical transmission, containment, PPE, and decontamination still sit between an autonomous protocol and host exposure. Automation can cut human splash and spill exposure while introducing new failure modes (mis-pipette, aerosolization, continuing when a human would stop). Practical implications: test autonomous loops on low-risk systems first; fold autonomous decision-making into biosafety risk assessments because the executed plan can drift after initial approval; define human-in-the-loop go/no-go criteria. Companion primaries include the EBRC 2023 white paper on engineering biology and AI security, DOE Genesis Mission autonomous laboratories, and NSF’s AI-programmable cloud labs network investment.
Closed-loop AI-directed experimentation has now been demonstrated in a benign optimisation task. A Ginkgo Bioworks and OpenAI preprint reports an autonomous lab pairing GPT-5 with an automated cloud laboratory to optimise cell-free protein synthesis, with iterative experimental design, execution, data capture, analysis, interpretation, and hypothesis generation handled by the system and human intervention largely limited to reagent and consumables preparation, loading, and unloading. The reported outcome was a 40% reduction in specific cost per gram of protein relative to the prior state of the art, with a 27% increase in titer (Smith et al., 2026, preprint).
Read that result precisely. It establishes that design-analyse-iterate loops can run with minimal human intervention on a well-characterised, non-hazardous production target where the assay is fast, cheap, and machine-readable. It does not establish that a general-purpose agent can autonomously execute a hazardous biological workflow, which requires pathogen handling, containment, materials access, and validation steps this system never touched. The governance-relevant claim is that the human checkpoint some oversight models assume is present can be engineered out of the loop, not that dangerous biology has been automated.
A separate bounded result (Fushimi et al., Nature Scientific Reports, 2025) shows an autonomous lab can improve E. coli growth for a glutamic-acid-producing strain yet fail to improve glutamic acid productivity under the same medium-optimization setup, a reminder that autonomy tracks objective complexity and data quality, not headline generality (DOI 10.1038/s41598-025-89069-y).
The Cloud Lab Ecosystem
What Are Cloud Labs?
Cloud laboratories are facilities where robotic systems execute biological or chemical experiments designed remotely by users. The user specifies what they want done - through a web interface, API, or specialized software - and automation handles the physical execution.
The model parallels cloud computing: instead of maintaining your own servers (wet lab), you rent access to shared infrastructure (cloud lab) and pay for what you use.
Core Capabilities:
- Liquid handling and sample preparation
- PCR, cloning, and molecular biology workflows
- Cell culture and fermentation
- Analytical chemistry and quality control
- High-throughput screening
Major Platforms
| Platform | Focus | Model |
|---|---|---|
| Emerald Cloud Lab | General life sciences | Full-service cloud lab |
| Culture Biosciences | Bioprocess optimization | Fermentation-as-a-service |
| Synthego | CRISPR workflows | Gene editing services |
| University biofoundries | Academic research | Institutional access |
These platforms serve legitimate and valuable purposes: enabling startups without wet lab space, improving reproducibility through standardized protocols, reducing costs through shared infrastructure, and expanding research access.
Publicly funded biofoundries are more than academic rows in a cloud-lab platform table. Vickers and Freemont argue that this synthetic biology infrastructure should be available to nations as part of independent bioengineering, biosecurity, and countermeasure response systems, including through coordinated Global Biofoundry Alliance networks (Vickers and Freemont, 2022). Imperial College London’s London Biofoundry (founded 2016; Global Biofoundry Alliance member from 2019) describes a COVID-19 pivot that repurposed high-throughput robotics for NHS diagnostic workflows, secured UKAS accreditation with North West London Pathology partners, fielded a reagent-agnostic open platform during global kit shortages, and supported nearly 2 million tests spanning staff, student, and asymptomatic community testing (Imperial College London / Freemont). Imperial story is institutional program narrative. Treat Vickers and Freemont Nature Communications as the peer framing for the national-infrastructure claim, and verify operational counts against primary program records before using them as procurement or legal facts.
How Cloud Labs Work in Practice
A typical workflow:
- Design: User specifies experiment through interface or API
- Review: Platform may screen for hazardous materials or protocols
- Scheduling: Experiment enters queue for robotic execution
- Execution: Automated systems perform procedures
- Results: Data and samples returned to user
The entire process can occur without the user ever entering a laboratory or handling biological materials directly.
The Biosecurity Gap
Traditional biosecurity frameworks rest on four assumptions that cloud labs disrupt:
Assumption 1: Physical Presence of Trained Personnel
Traditional model: A human with laboratory training physically performs experiments. This person has been vetted, trained, and is accountable (their institution knows who they are).
Cloud lab disruption: The user may never enter a laboratory. Their only interaction is through a computer interface. The verification that occurs is customer onboarding, not laboratory training.
Assumption 2: Institutional Biosafety Review
Traditional model: Experiments involving biological materials go through institutional biosafety committees (IBCs). The IBC reviews protocols, ensures compliance with regulations, and provides oversight.
Cloud lab disruption: Who is the “institution” when a user in Country A submits an experiment to a robotic lab in Country B? The cloud lab company may have its own review processes, but these may not be equivalent to traditional IBC oversight and may vary significantly between providers.
Assumption 3: Physical Access Controls
Traditional model: Laboratories have physical security - badge access, locked doors, restricted areas. These controls limit who can enter and what they can access.
Cloud lab disruption: Access is now mediated through software credentials - usernames, passwords, API keys. These can be shared, stolen, or misrepresented more easily than physical presence.
Assumption 4: Human Observation
Traditional model: Lab colleagues might notice suspicious activity. A lab manager might question an unusual request. Tacit social controls exist because experiments happen around other people.
Cloud lab disruption: When experiments are executed by robots and monitored remotely, there are fewer opportunities for human observation to flag concerning patterns. The “social” aspect of laboratory oversight disappears.
Some biosecurity analysts describe cloud labs as a potential “loophole” - a path around traditional controls rather than through them.
This framing can be misleading. Providers describe screening and verification measures, but coverage and rigor vary and are not uniformly transparent. The concern is that existing regulations were not designed with this model in mind, creating ambiguity about requirements and accountability.
A RAND Corporation analysis identifies 15 cloud lab organizations and describes the sector as creating “potential loopholes in existing chemical and biological weapons nonproliferation frameworks,” precisely because users can outsource experimental work to facilities not clearly covered by existing regulations. Similarly, SIPRI’s 2025 analysis identifies cloud labs as “new actors” that sit uneasily within traditional export control regimes, with relevant activities mediated through digital interfaces rather than on-site human operation.
A further development extends the concern: self-driving laboratories (SDLs) move beyond democratizing access to laboratory infrastructure and begin democratizing experimental decision-making itself. In SDL configurations, route planning, condition selection, and iterative optimization shift from human chemists to AI systems, reducing reliance on the tacit expertise that has historically bounded who could conduct advanced chemical and biological work. The OPCW Scientific Advisory Board’s TWG on AI (SAB/REP/1/26, 2026) notes that this “deskilling” of complex tasks, combined with remote execution and digitally mediated access, may challenge traditional approaches to monitoring and verification compliance.
The risk is not that cloud labs are uncontrolled, but that controls may be inconsistent across providers, and the regulatory framework has not caught up to the technology.
AI + Cloud Lab Convergence: The Concerning Scenario
The biosecurity community has increasingly focused on what happens when AI capabilities gain access to automated laboratory execution.
The Scenario
Consider a progression:
- Today: An AI can suggest experiments, provide troubleshooting guidance, and design biological sequences
- Near-future: An AI agent with cloud lab API access could iterate experiments autonomously - all without human intervention
Component capabilities are demonstrated: AI agents conducting chemistry experiments have already been published. Full integration with cloud biology laboratories remains theoretical. The technical pathway exists, but no published demonstration of autonomous AI-driven biological experimentation via cloud labs has occurred.
What Would Need to Happen
For the concerning scenario to materialize:
- API Access: The AI would need credentials to submit work to a cloud lab
- Protocol Capability: The AI would need to generate valid experimental protocols
- Iteration Capability: The AI would need to interpret results and design follow-up experiments
- Minimal Human Oversight: The human “owner” would need to be absent, negligent, or complicit
Current cloud labs require human accounts and some level of verification. But as AI agents become more capable, the distinction between “human using AI assistance” and “AI acting autonomously” may become harder to detect.
In LLMs and Information Hazards, we described the current era as “Google Plus” - LLMs are essentially better search engines that synthesize information more efficiently.
The day an LLM is directly connected to a cloud laboratory - a robotic lab controllable over the internet - is the day the tacit knowledge barrier may begin to dissolve.
This is not a prediction that disaster follows immediately. But it is a qualitative shift in the threat landscape that warrants proactive governance rather than reactive response.
Near-Term Realistic Concerns
Before the fully autonomous scenario, more immediate concerns include:
Novice uplift: Cloud labs could enable people with theoretical knowledge but no lab skills to conduct experiments, reducing the practical barrier to entry.
Distributed bad actors: A threat actor could submit seemingly innocuous work from multiple accounts, with concerning elements only apparent in aggregate.
Screening evasion: Novel AI-designed sequences (per AI-Enabled Pathogen Design) that evade traditional screening could be synthesized through cloud lab platforms if those platforms rely on the same screening approaches as DNA synthesis providers.
Demonstrated (supported by published evidence):
- Cloud laboratories are operational and commercially available (Emerald Cloud Lab and others)
- Remote users can design and execute experiments without physical lab access
- Providers publicly describe customer verification and protocol screening, but controls vary and independent testing is limited
- Automation remains imperfect - complex biology still requires human troubleshooting
- No publicly documented cloud-lab biosecurity incident was identified in the sources reviewed
Theoretical (plausible but not yet demonstrated):
- AI agents autonomously using cloud labs to execute dangerous experiments
- Cloud labs substantially lowering barriers for novice threat actors
- Distributed attacks across multiple accounts evading aggregate detection
- AI-designed sequences evading cloud lab screening systems
Unknown (insufficient evidence to assess):
- How quickly automation will mature to reduce human oversight requirements
- Whether existing screening approaches are adequate for emerging threats
- The effectiveness of current Know Your Customer protocols against sophisticated adversaries
- How governance frameworks will adapt as capabilities advance
The window for proactive governance exists now - before capabilities become fully integrated. Current limitations provide time to establish appropriate safeguards.
Current Governance and Gaps
How Cloud Labs Screen Today
Responsible cloud lab providers implement various controls:
Customer verification:
- Identity verification during account creation
- Institutional affiliation checks for certain capabilities
- Export control compliance (ITAR, EAR) where applicable
Protocol review:
- Automated screening for select agents and regulated materials
- Human review for flagged protocols
- Restrictions on certain experiment types
Activity monitoring:
- Logging of all submitted work
- Anomaly detection for unusual patterns
- Audit trails for regulatory compliance
Material controls:
- DNA synthesis screening (for labs that synthesize DNA)
- Sourcing restrictions for biological materials
- No handling of select agents or tier 1 pathogens
Regulatory Ambiguity
The challenge is that existing regulations were designed for traditional laboratory settings:
FSAP (Federal Select Agent Program): Regulates possession and use of select agents but focuses on physical facilities and registered entities. Cloud labs that do not handle select agents may fall outside this framework.
IBC Requirements: NIH Guidelines require IBC review for certain research, but applicability to cloud lab work may be unclear, especially for non-NIH-funded users.
Export Controls: ITAR and EAR apply to certain biological materials and technologies, but enforcement for remote access services across jurisdictions is complex.
State and Local Regulations: Laboratory regulations vary by jurisdiction, but cloud labs may operate in different jurisdictions than their users.
The cloud-lab industry is developing voluntary practices, including:
- Customer screening procedures
- Protocol review processes
- Information sharing on concerning requests
- Engagement with government on regulatory frameworks
Whether voluntary measures are sufficient, or whether formal regulation is needed, remains an open question. RAND has called for a “Cloud Lab Security Consortium” - modeled on the International Gene Synthesis Consortium (IGSC) - to standardize screening across providers and share threat intelligence.
Proposed Safeguards
Several frameworks have been proposed for managing cloud lab biosecurity risks:
Enhanced Screening
Protocol-level review: Screen not just DNA sequences but experimental protocols for concerning combinations or outcomes.
Aggregate monitoring: Look for patterns across experiments from the same user that might indicate stepwise progression toward concerning capabilities.
AI-assisted screening: Use AI to identify concerning experimental designs, not just sequence matches.
Human-in-the-Loop Requirements
Verification for high-risk work: Require video verification, institutional endorsement, or other enhanced authentication for experiments above certain risk thresholds.
Mandatory review points: Certain protocol types trigger human review before execution proceeds.
Rate limiting: Constrain how quickly new users can access sensitive capabilities.
Technical Controls
API restrictions: Limit automated submissions for new accounts or sensitive experiment types.
Anomaly detection: Flag unusual patterns of usage for human review.
Audit trails: Maintain detailed logs for regulatory and law enforcement access.
Governance Frameworks
Clear regulatory jurisdiction: Establish which agency has authority over cloud lab biosecurity.
Provider standards: Develop industry-wide minimum screening requirements.
International coordination: Work with international partners to prevent jurisdiction shopping.
“Know Your Customer” (KYC) Standards
Just as banks must verify client identity to prevent money laundering, cloud labs must verify the identity and intent of their users. This parallel to financial sector KYC requirements suggests:
- Verified institutional affiliation
- No anonymous accounts
- Clear audit trails for all transactions
- Shared “no-fly lists” of banned users across providers
If you are reviewing a cloud lab platform (as a funder, regulator, or potential user), ask:
1. Customer Verification
- What identity verification occurs during onboarding?
- Is institutional affiliation required for certain capabilities?
- How are export control requirements addressed?
2. Protocol Screening
- What screening occurs before experiment execution?
- Is DNA synthesis screening applied?
- Are protocols reviewed for dual-use concerns?
3. Activity Monitoring
- What logs are maintained?
- Is anomaly detection implemented?
- How are concerning patterns reported?
4. Governance
- Which regulations apply to this platform?
- What voluntary standards are followed?
- How is the platform engaging with regulators?
Benefits and Legitimate Uses
It is important to emphasize: cloud labs are not primarily a biosecurity problem. They offer genuine value for legitimate science.
Why Cloud Labs Matter for Science
Access expansion: Researchers at institutions without wet lab space can conduct experimental work.
Reproducibility: Standardized, robotic execution can reduce variability between experiments.
Cost reduction: Shared infrastructure reduces the capital investment needed for experimental research.
Speed: Parallel execution and optimized scheduling can accelerate research timelines.
Training: Some platforms are used for education, allowing students to “conduct” experiments remotely.
Who Uses Cloud Labs Legitimately
- Biotech startups without dedicated lab space
- Academic researchers accessing specialized equipment
- Pharmaceutical companies for high-throughput screening
- Synthetic biology teams for standardized DNA assembly
- Educational institutions for remote learning
Balancing Access and Security
The goal should not be to eliminate cloud labs but to implement risk-proportionate controls. The framework should:
- Preserve access for legitimate research
- Screen effectively for concerning requests
- Monitor actively for misuse patterns
- Respond rapidly when concerns arise
This is analogous to the approach for DNA synthesis: not to ban the technology, but to implement screening that catches concerning orders while allowing the vast majority of legitimate work to proceed.
International Dimensions
Cloud labs operate globally, creating coordination challenges:
Multi-jurisdictional operation: A company may be headquartered in one country, operate facilities in another, and serve customers worldwide.
Regulatory arbitrage: If one jurisdiction implements strict controls while another does not, users might shop for the least restrictive option.
Export control complexity: The remote nature of cloud labs complicates traditional export control enforcement.
Dual-use considerations: The Biological Weapons Convention prohibits development of biological weapons but relies on national implementation that may not address cloud lab access.
Effective governance will require international coordination - likely through existing frameworks like the Australia Group or new mechanisms specifically addressing cloud biology.
Reality Check: The “Broken Robot” Factor
A necessary dose of skepticism is warranted here.
If you read discussions among biotech engineers who work with these systems, a common theme emerges: automation is hard.
The failure rate: Liquid handling robots jam. Tips fall off. Reagents expire. Incubators drift. Contamination happens.
The tacit barrier persists: Even with robots, biology requires troubleshooting. “The cells look weird today” is an observation a robot cannot easily make or act upon. The experienced lab scientist who notices something is off remains essential.
The iteration problem: In practice, most cloud lab experiments require human interpretation between iterations. Fully autonomous closed-loop systems work for narrow optimization problems, but complex biology still requires human judgment.
Conclusion: While the security risks warrant proactive governance, no published evidence reviewed here shows cloud labs autonomously generating pandemic agents. Wet-lab automation remains constrained by reliability and troubleshooting requirements, creating an opportunity to improve controls before more autonomous workflows become routine.
Autonomous AI Agents: An Emerging Attack Surface
Autonomous AI agents represent a distinct and expanding threat model that extends beyond cloud laboratories. These software systems combine large language models with the ability to browse the web, execute code, call APIs, and chain multiple tools together to accomplish complex goals with minimal human oversight. When agents gain API access to laboratory infrastructure, they can design experiments, submit protocols, interpret results, and iterate autonomously.
Open-source agent frameworks such as OpenClaw illustrate how agent tooling can spread quickly. Popularity metrics do not establish laboratory capability, but the combination of agent tools and cloud-laboratory APIs is a reason to assess access controls before autonomous workflows become routine.
For comprehensive treatment of agent-specific risks, including skill supply-chain vulnerabilities, multi-agent coordination, governance frameworks like “least agency,” and case studies of emerging agent ecosystems, see Autonomous AI Agents in Laboratory Contexts.
Future Trajectory
Increasing Automation
The trend is toward greater automation and AI integration:
- More experimental capabilities becoming automated
- Better APIs enabling programmatic access
- AI systems increasingly capable of experimental design
- Integration of design-build-test-learn cycles
Governance Keeping Pace?
The question is whether governance can keep pace with these capabilities. The window for proactive action may be limited:
Today: Cloud labs are a niche technology used primarily by sophisticated researchers. Controls can be implemented before widespread adoption.
Tomorrow: Cloud labs become a standard part of the research infrastructure. Implementing controls becomes harder as incumbents resist and alternatives proliferate.
The biosecurity community has an opportunity to shape cloud lab governance before the technology becomes ubiquitous.
Learning from previous technologies (DNA synthesis, gain-of-function research), it is easier to establish norms and requirements early than to retrofit them onto mature industries. The lesson is not to wait until a concerning incident occurs.
What are cloud laboratories and how do they work?
Cloud laboratories are facilities where users design experiments remotely through web interfaces or APIs, and robotic systems execute the physical work automatically. Users specify their experimental protocols, the platform screens for hazardous materials, and automated systems perform liquid handling, PCR, cell culture, and other procedures without requiring physical lab access. Results are returned electronically.
How do cloud labs differ from traditional laboratory biosecurity?
Traditional biosecurity assumes trained personnel physically present in laboratories, institutional biosafety committee oversight, physical access controls like badge systems, and human observation of suspicious activities. Cloud labs disrupt all four assumptions: users never enter the lab, institutional oversight may be ambiguous across jurisdictions, access is controlled through software credentials that can be shared or stolen, and robotic execution eliminates social observation of concerning work.
What is the AI convergence risk with cloud laboratories?
The convergence concern arises when AI systems capable of designing experiments gain API access to cloud labs for autonomous execution. AI could potentially design protocols, submit them remotely, interpret results, and iterate new experiments without human intervention. This would dissolve the “tacit knowledge barrier” that currently limits who can conduct biological experiments, potentially enabling novices to execute dangerous work.
Are cloud labs currently being misused for biosecurity threats?
No publicly documented cloud-lab biosecurity incident was identified in the sources reviewed. Providers describe customer verification, protocol screening, and activity monitoring, but practices vary. Wet-lab automation remains imperfect: robots jam, reagents fail, and complex biology still requires human troubleshooting. The fully autonomous threat scenario is not established in the published record, providing an opportunity to strengthen governance before capabilities mature.
This chapter addresses cloud laboratories and automated biology. For handbook-wide priorities, see the Executive Summary. For the autonomous agent threat model that could operate this infrastructure, see Autonomous AI Agents in Laboratory Contexts. For earlier context, see AI as a Biosecurity Risk Amplifier, LLMs and Information Hazards, AI-Enabled Pathogen Design, AI for Biosecurity Defense, and Digital Biosurveillance.