Attack types
Types of data poisoning attack, by where they strike
Poisoning can strike at pre-training, fine-tuning or embedding, and the model file itself can be tampered with. Each route needs a different control.
Last reviewed against Anthropic, UK AI Security Institute and arXiv. Sources
This page sorts the routes the sources describe. It does not rank them: none of the sources gives figures for how often each is used.
By stage
Where the data is altered
OWASP separates the stages of a model's life at which data can be manipulated [4]. The routes differ in who controls the data.
Pre-training data
General data gathered at large scale. Researchers showed that crawled web datasets can be poisoned through expired domains and through timed edits to crowd-edited pages [8].
Fine-tuning data
Task-specific data, often smaller and closer to home. The 2025 preprint reported similar dynamics when poisoning during fine-tuning [3].
Embedding data
Text converted into vectors for retrieval. OWASP lists embedding data alongside the other two stages as open to manipulation [4].
Two web-scale attacks
Split-view and frontrunning
A 2023 paper by researchers including Nicholas Carlini describes two attacks on datasets built from web content [8].
The file itself
Malicious model files
Poisoned data is not the only risk with shared models. OWASP warns that models on public repositories can carry malicious code, such as payloads hidden in pickle files that run when the model is loaded [4].
Check these controls against each type:
- For pre-training data, know the source and keep a verified copy
- For fine-tuning data, restrict who can add to it
- For embedding stores, control what content can be ingested
- For downloaded models, treat the file as software and scan it before loading
Related risk
Poisoning by injection
OWASP's scenarios include weak filtering that lets prompt injection insert misleading data into training material [4]. That is why input filtering and data controls belong together.
Where next
Putting numbers on it
The 250 document study shows how small an attack can be, and the prevention guide covers the controls.
Keep reading
Read next
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 guide
Web datasets
How researchers showed that datasets crawled from the web can be poisoned cheaply, through expired domains and timed edits.
Read the guide
Medical case
A reported Nature Medicine study on poisoning medical language models with 0.001% misinformation, and why benchmarks did not catch it.
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.