DNA evidence has anchored criminal convictions for decades, trusted because judges and juries assume a machine readout never lies about the physical sample inside it. That assumption has just been proved that completely wrong. Researchers used AI-generated code to alter raw forensic DNA files. The laboratory analysis software was entirely unaware of the manipulation.
The demonstration, involving Anthropic’s Claude to generate the necessary code, took a systems engineer at Forensic Bioinformatics named Nathan Adams approximately 45 minutes to execute successfully. The modified file combined scans from two separate DNA profiles into a new file that appeared unaltered and carried a date stamp reaching back to 2015. The lab analysis software raised no warnings.
Thermo Fisher Scientific, which manufactures the affected equipment, privately acknowledged the flaw in July 2026. It has assigned it a High severity rating under CVE-2026-17583 with a CVSS v4.0 score of 8.2.
What Tampering With Digital DNA Files Enables
The vulnerability is inside the digital file system. The physical evidence itself is never touched.
Applied Biosystems human-identification systems, used in most US crime labs, produce .fsa and .hid files without tamper-evident digital signatures. Any changes made to these files pass through without raising flags. An attacker with access to those files before they are loaded into analysis software can alter them, combining profiles, removing a suspect’s DNA or inserting an innocent person’s profile. The physical swab inside the evidence bag remains completely untouched. Only the digital data shown to analysts and presented in court gets altered.
The researchers found no way to catch previous file tampering. That detail changes the equation. Investigators can track down a forgery that leaves clues behind. A forgery that leaves no footprint places the burden of proof in a whole new light. For cases where digital files have already been checked against the physical sample and accepted by courts, there’s currently no known forensic technique to determine retrospectively whether the data was modified.
More from Artificial Intelligence
- Is Rapid AI Price Deflation A Gift Or An Existential Threat To SaaS?
- Airlock Digital Unveils Agentic AI Control And Governance To Extend Preventative Endpoint Security
- Why Are VCs Pulling Back From Open-Weight AI Startups?
- Why Is Google DeepMind Calling Gemini Robotics 2 A Step Toward Physical AGI?
- What Is Retrieval-Augmented Generation?
- You Can Now Report AI Slop On LinkedIn – Assuming You Can Spot It
- Anthropic’s Three AI Breaches Are A Wake-Up Call For AI Safety – Here’s Why
- Is AI Being Blamed For A Problem Humans Created?
The Threat To Decades Of Historical Evidence
Thermo Fisher has issued patches that add digital signatures to verify file integrity going forward. Three older product lines have reached end-of-life status and won’t receive security fixes. The company recommends strict access restrictions and encrypted storage for laboratories unable to patch. This flaw creates huge historical exposure because it impacts digital files produced by crime lab equipment back to 1995, putting thirty years of case files at risk.
New files generated after patching can be verified. Historical files can’t be retroactively secured, and there is no known method to detect whether any existing file has been modified. For defence teams in ongoing cases, this creates grounds to challenge the integrity of digital DNA files and seek audits, re-testing or enhanced provenance checks. Courts have yet to deal with the fallout of using a forensic file format that could be modified without a trace for decades.
How Has AI Changed The Reality Of This Attack?
The technique isn’t new in principle. Forensic file manipulation has been theoretically possible wherever digital evidence lacks cryptographic protection. What AI has changed is the barrier to entry. A targeted attack on a specialised forensic file format previously required deep knowledge of the format’s internal structure, the analysis software’s validation logic and the specific instrumentation involved. With an AI coding assistant, a systems engineer without a background in forensic bioinformatics produced a working exploit in under an hour.
This pattern makes AI-assisted attacks completely different from older threats. An attacker no longer needs years of specialised knowledge to carry out an advanced exploit. They only need access to a model capable of generating the necessary steps. Lower costs, improved capabilities and wider availability of AI coding tools mean that specialised attacks are no longer restricted to elite experts. Anyone with basic technical skills and thirty minutes can now target high-stakes systems.
Together with recent Anthropic containment breaches, this DNA evidence case highlights the true nature of practical AI safety risks. Forget science fiction scenarios about rogue superintelligence. This is about documented proof that AI tools allow ordinary users to launch once impractical attacks quickly and cheaply.
