Risks
The risks data poisoning creates, and the evidence
Poisoned data can leave a system that passes its tests yet misbehaves when a hidden trigger appears. The evidence so far comes from research, not from a named incident.
Last reviewed against Anthropic, UK AI Security Institute and arXiv. Sources
Hidden behaviour
A backdoor that stays quiet
What this means for assurance
A pass on routine tests is not proof that the training data was clean. Treat benchmark results as one input, not a clearance.
Low cost
Less control needed than assumed
The 2025 study found that the count of poisoned documents mattered, not their share of the data [1]. For the 13 billion parameter model, 250 documents were about 0.00016% of training tokens [1]. An earlier paper said about 0.01% of two web-scale image datasets could have been poisoned for about $60 [8].
Read the limits too
The study authors note that attackers still face practical barriers, such as getting poisoned content into a training set [1].
Routes in
Where tampered data enters
OWASP lists the points where data can be altered: pre-training, fine-tuning and embedding [4]. Its example scenarios include falsified documents supplied for training and weak filtering that lets injected content through [4].
Check these as well:
- Models downloaded from public repositories, which OWASP says can carry malicious code that runs on loading [4]
- Datasets crawled from the web or built from snapshots of crowd-edited pages [8]
- Third-party data vendors whose collection methods you have not checked [4]
- Retrieval and embedding stores fed from outside content [4]
Real consequences
Wrong outputs in serious settings
The medical study reported that replacing about 0.001% of training tokens with misinformation produced models that often gave erroneous medical statements [9]. For a UK organisation the exposure depends on what the system decides or advises: the more consequential the output, the more the integrity of the data matters. Read the medical study guide for detail.
Next step
Reduce what you can
The prevention guide and the detection guide set out controls drawn from OWASP and the NCSC.
Key facts
Data Poisoning in brief
- What it is
- Manipulating pre-training, fine-tuning or embedding data to add weaknesses, backdoors or bias[4]
- Documents needed
- 250 poisoned documents backdoored every model tested[1]
- Models tested
- 600 million to 13 billion parameters[3]
- Share of training data
- About 0.00016% for the 13 billion parameter model[1]
- Web-scale datasets
- About 0.01% of LAION-400M or COYO-700M poisoned for about $60[8]
- Medical study
- 0.001% of training tokens replaced with misinformation[9]
Keep reading
Read next
What it is
A plain explanation of data poisoning: how tampered training data changes what an AI system does, and why backdoors are hard to find.
Read the guide
Attack types
The main ways training, fine-tuning and retrieval data can be poisoned, and how poisoned model files differ, with sources.
Read the guide
The 250 study
What the Anthropic, UK AI Security Institute and Alan Turing Institute study found about poisoning language models, and what it does not show.
Read the guideHow CyPro can help
Need a hand planning for this?
Understanding the change is one job. Working out what it means for your systems, your suppliers and your roadmap is several more. CyPro's consultants can take any of them on.
Fractional CISO for the board
A senior security lead who advises the board on what to fix first, what to accept and what to tell customers, and owns the decision record.
More on cypro.co.uk
Hardening and segmentation
Our consultants put the controls in place that limit the damage when something is missed: firewall rules, network isolation and tighter admin access.
More on cypro.co.uk
Penetration testing
A focused test of your systems and the controls around them, so you know they hold before an attacker finds out for you.
More on cypro.co.uk
Vulnerability scanning
Find every internet-facing system you run, including the forgotten ones, and see which are still on a vulnerable build.
See the scanning service
24/7 security monitoring
Our SOC analysts watch your systems day and night for the signs of an attack in progress: unexpected processes, new outbound connections, logins that do not fit.
More on cypro.co.uk
Incident response
If you may already be compromised, our responders contain it, preserve the evidence, find out what was touched and get you running again safely.
More on cypro.co.uk
Plan ahead
Want a plan for your organisation?
Defending against data poisoning starts with knowing which AI systems you run, where their data comes from and who owns the risk. CyPro can help set the governance, risk assessment and controls around that.