Biosecurity in cyberspace

Cyberbiosecurity is the seam where digital control meets biological capability. Two mechanisms deserve explanation, because both have limits that are structural rather than a matter of effort.

Screening an order is a similarity search

Synthetic DNA is ordered as a text file. The defence is that providers screen orders before they synthesise: against the customer, and against the sequence itself. The industry’s Harmonized Screening Protocol, run through the International Gene Synthesis Consortium, and the US HHS Screening Framework Guidance for Providers of Synthetic Nucleic Acids set the pattern — flag orders whose sequence resembles agents on the Select Agent and export-control lists, then adjudicate the hit by hand.

The mechanism is homology search, and its weaknesses follow directly. It needs a length to search over: the HHS framework’s baseline applies to sequences of 200 nucleotides or more, so shorter fragments ordered separately and assembled by the customer pass beneath it. It matches against a list, so an order can be split across several providers, each of whom sees a fragment that resembles nothing. And it matches nucleotide similarity, while what matters is protein function: recoding a gene with synonymous codons, or redesigning a protein to keep its fold and activity while diverging in sequence, can drop identity below any workable threshold without changing what the molecule does. Raise the threshold’s sensitivity and the false-positive load lands on human reviewers, which is the real operating constraint.

Two developments push on this. Benchtop synthesisers move synthesis out of the screened supply chain entirely, so the screen has to move into the device and its firmware — which makes device attestation, not sequence matching, the control point. And protein design models can propose functional sequences with no meaningful homology to anything catalogued, which strains the founding assumption that dangerous means recognisable. How well screening generalises to designed sequences is not settled, and claims in either direction should be treated as open.

The threat to a laboratory is integrity, not theft

Security thinking defaults to confidentiality. In a laboratory the more consequential loss is silent alteration: a modified chromatogram, a shifted setpoint on a fermenter’s control system, a swapped sample identifier in a LIMS. Nothing is exfiltrated, so no confidentiality control fires, and the result is a real experiment that reports a false answer — with the falsification propagating into every downstream decision that trusted it. This is why regulated environments insist on attributable, contemporaneous, unalterable audit trails; the audit trail is a detection mechanism as much as a compliance artefact.

The instruments themselves are the weak point, for a reason that resists the usual fix. Sequencers, mass spectrometers and bioreactor controllers are often driven by vendor software pinned to an old operating system, cannot be patched without the vendor’s release, and — in a GxP environment — cannot be changed at all without revalidating the system. Patch promptly and you invalidate qualification; leave it unpatched and you run a known-vulnerable host on the network. The workable answer is compensating controls, principally network segmentation that isolates instrument networks from general IT, rather than the patching cadence used everywhere else.

That instrument boundary is genuinely bidirectional. A 2017 study presented at USENIX Security showed that DNA could be synthesised to encode an exploit triggered when a sequencer’s downstream software processed the reads — a proof of concept that required deliberately weakened software, and useful mainly as evidence that biological material is an input channel like any other.