ResearchPapers, evidence & method
Google DeepMind watermarks AI-designed proteins as new studies test agent swarms and AI-run labs
Google DeepMind reports that watermarked AI-designed proteins performed as well as unmarked ones in wet-lab tests on three targets. Separate work finds that agent swarms trade efficiency for speed, and that the best model passes under half of real lab tasks.
AT Pulse research deskOctober 6, 2026 · 5 min read · AI-assisted analysisSynthID Bio adds a watermark in two ways. One method steers amino-acid choices in protein sequences. The other shifts atomic coordinates in predicted structures, with the mark built into AlphaFold 3's diffusion weights. DeepMind reports that, across three targets in wet-lab tests, watermarked binders matched unwatermarked ones on hit rate, binding affinity and sequence diversity. These results come from DeepMind itself. Outside experts spoke well of the work but have not repeated the tests, and DeepMind says resistance to deliberate tampering still needs work. [2] [8]
Toby Ord read data off charts in OpenAI's GPT-5.6 launch post and estimated a swarm-scaling exponent of about 0.48 to 0.68 across three benchmarks. On those numbers, ten times more agents gives roughly a third to a half of the gain that ten times more tokens gives a single agent. A four-agent swarm finishes about twice as fast for about twice the compute. Wenhao Chai reports a much higher exponent, 0.88 to 0.93, for recursive swarms working through task graphs. That suggests the way a swarm is organised changes the result. [3] [7]
C5R's SciUniverse benchmark has 92 tasks across chemistry, biology and materials science. Models write experiments as code, which runs on lab instruments with help from human operators. The top entry, Claude Fable 5.1 at its highest effort setting, passed 45.3% of tasks at about $40 per task. Other frontier models scored much lower. The scores come from the benchmark's builder, rest on only a few noisy real-world runs, and include some simulated tasks. [4] [5]
A position paper by DeepMind authors argues that physical lab resources, not ideas, will limit AI scientists. It proposes a market to decide which experiments get run: a record of when each idea was proposed, forecasting markets that price ideas before testing, licensing of ideas to labs, and royalties once an idea is validated. Jack Clark presents these items as early warning signs for automated science. The paper is an argument and offers no empirical evidence. [6] [1]
results from DeepMind and C5R suggest that tagging AI-designed proteins can work and that AI-run labs still need human help.
Read the full assessment
Implication: DNA synthesis screeners could gain a useful signal if model providers adopt the watermark, while agent swarms and lab automation look like costly, supervised speed-ups.
Executive brief
Google DeepMind reports that it can hide a detectable watermark inside AI-designed proteins without hurting how they work. In wet-lab tests on three targets, watermarked binders matched unwatermarked ones on hit rate, binding affinity and sequence diversity (Google DeepMind). These results come from the vendor and have not been independently replicated. Jack Clark's Import AI 475 puts this alongside three other items. Toby Ord estimates that agent swarms scale worse than single agents given the same tokens, but finish faster. A new lab benchmark finds the best frontier model passes 45.3% of real lab tasks. A DeepMind position paper proposes markets to decide which AI-generated experiments get run.
What changed and event timeline
Ord quantifies swarm scaling
Using charts from OpenAI's GPT-5.6 launch post, Ord estimates λ between 0.48 and 0.68 across three benchmarks. At that rate, 10x more agents yields roughly 3–5x the gain that 10x more tokens gives a single agent ().
C5R launches SciUniverse
The startup unveiled Facility-0, a lab run by AI models with human operators, and a 92-task benchmark spanning chemistry, biology and materials science (;).
DeepMind authors propose a science economy
A position paper by Tomasev, Franklin, Kasirzadeh and others argues that physical resources, not ideas, will be the bottleneck for AI scientists ().
Swarm results that disagree with Ord
Wenhao Chai reports λ of 0.88–0.93 for recursive swarms working through task graphs, well above Ord's range ().
SynthID Bio announced
DeepMind released watermarking for protein sequences and predicted structures, with open code and data (). TNW dates its coverage to October 1 and mentions a Nature publication ().
Import AI 475 published
Clark argues that 2026 is producing "warning shots" for automated science, just as it has for automated AI R&D ().
Capabilities and access
- SynthID Bio: one method shapes amino-acid choices in sequences; another adjusts atomic coordinates in predicted 3D structures, with the watermark built into AlphaFold 3's diffusion weights.
- SciUniverse leaderboard (Pass@1): Claude Fable 5.1 (xhigh) 45.3% at $40.61 per task; Claude Opus 5 30.5%; Grok 4.6 26.2%; Gemini 3.8 Flash 14.6%; GPT-5.6 Sol 9.4% (C5R).
Read the full section
- SynthID Bio: one method shapes amino-acid choices in sequences; another adjusts atomic coordinates in predicted 3D structures, with the watermark built into AlphaFold 3's diffusion weights. DeepMind is open-sourcing the code and in vitro data and releasing weights to researchers (DeepMind).
- SciUniverse leaderboard (Pass@1): Claude Fable 5.1 (xhigh) 45.3% at $40.61 per task; Claude Opus 5 30.5%; Grok 4.6 26.2%; Gemini 3.8 Flash 14.6%; GPT-5.6 Sol 9.4% (C5R). The source page does not say whether the tasks are public.
Technical analysis for researchers and developers
- Ord's method: he read data points off OpenAI's published charts for 1-, 4- and 16-agent swarms on BrowseComp, SEC-Bench Pro and Terminal-Bench, then interpolated on log plots.
- SciUniverse design: 17 task families, each weighted equally. Some tasks are simulated (C5R).
- SynthID Bio detection: uses a cryptographic key and statistical thresholds (TNW).
Read the full section
- Ord's method: he read data points off OpenAI's published charts for 1-, 4- and 16-agent swarms on BrowseComp, SEC-Bench Pro and Terminal-Bench, then interpolated on log plots. Claude Opus 5 ran the regressions. λ was 0.68, 0.57 and 0.48 respectively, and a 4-agent swarm reached the same result about 2x faster for about 2x the compute (Ord). These are second-hand, chart-derived estimates.
- SciUniverse design: 17 task families, each weighted equally. Models write experiments as code, which is turned into instrument commands and instructions for human operators. Some tasks are simulated (C5R).
- SynthID Bio detection: uses a cryptographic key and statistical thresholds (TNW).
Claims and evidence
- Watermarking leaves protein function intact on three targets
- SciUniverse scores
- λ ≈ 0.5–0.7
Read the full section
| Claim | Status |
| Watermarking leaves protein function intact on three targets | Vendor-reported (DeepMind); outside experts quoted favourably but did not replicate (TNW) |
| SciUniverse scores | Reported by the benchmark builder; no independent replication found (C5R) |
| λ ≈ 0.5–0.7 | Independent analysis of vendor charts (Ord); disputed by Chai |
| Labs will be bottlenecked by physical resources | Argument, not evidence (arXiv) |
Context and prior work
- SynthID started as a watermark for AI-generated text and media.
- Ord ties λ to economists' "stepping on toes" effect, where adding people to a team yields diminishing returns (Ord).
- The science-economy paper has four parts: proof of when an idea was proposed, forecasting markets that price ideas before they are tested, licensing ideas to labs that run experiments, and royalties paid once an idea is validated (Import AI).
Read the full section
- SynthID started as a watermark for AI-generated text and media. The bio version was built with Stanford's Hie lab and the Arc Institute, alongside tools such as ProteinMPNN, Evo 2 and AlphaFold 3 (DeepMind).
- Ord ties λ to economists' "stepping on toes" effect, where adding people to a team yields diminishing returns (Ord).
- The science-economy paper has four parts: proof of when an idea was proposed, forecasting markets that price ideas before they are tested, licensing ideas to labs that run experiments, and royalties paid once an idea is validated (Import AI).
Limitations, safety and contested findings
- SynthID Bio: DeepMind says resistance to deliberate tampering still needs work (DeepMind).
- SciUniverse: few real-world runs make results noisy (C5R). Humans are in the loop, so this is not fully autonomous science (Runtime Wire).
- Swarm λ: Ord reports 0.48–0.68; Chai reports 0.88–0.93 in a different task-graph setup.
Read the full section
- SynthID Bio: DeepMind says resistance to deliberate tampering still needs work (DeepMind). TNW lists further weak points: very short proteins carry too little signal, fusing marked and unmarked proteins dilutes the mark, security depends on key management, and most protein design tools don't use ProteinMPNN (TNW).
- SciUniverse: few real-world runs make results noisy (C5R). Humans are in the loop, so this is not fully autonomous science (Runtime Wire). Model names don't match across sources: Import AI's "GPT-5 Astra" appears as "GPT-6 Astra" on C5R's page.
- Swarm λ: Ord reports 0.48–0.68; Chai reports 0.88–0.93 in a different task-graph setup.
Business and practitioner implications
- Agent builders: on Ord's numbers, swarms buy speed, not efficiency. Expect roughly double the token cost for a 4-agent speedup.
- Biotech and DNA synthesis firms: watermarks could help screeners decide which sequences need closer review, according to Twist Bioscience's James Diggans (TNW).
- R&D leaders: lab-automation pass rates below 50% at about $40 per task point to human-supervised pilots, not unattended labs (C5R).
Read the full section
- Agent builders: on Ord's numbers, swarms buy speed, not efficiency. Expect roughly double the token cost for a 4-agent speedup. Chai's higher λ suggests the organisational design of the swarm matters (Chai).
- Biotech and DNA synthesis firms: watermarks could help screeners decide which sequences need closer review, according to Twist Bioscience's James Diggans (TNW). They only work if model providers adopt them.
- R&D leaders: lab-automation pass rates below 50% at about $40 per task point to human-supervised pilots, not unattended labs (C5R).
Sources
Read the full section
- Import AI 475 (Jack Clark)
- Swarm Scaling (Toby Ord)
- Predictable Swarm Scaling (Wenhao Chai)
- SynthID Bio (Google DeepMind)
- Google DeepMind's watermarked AI proteins still work in the lab (The Next Web)
- SciUniverse (C5R)
- C5R is building physical labs for AI models (Runtime Wire)
- Agentic Economies for Autonomous Scientific Discovery (arXiv)