Your AI assistant just pulled a base image
AI coding tools write the Dockerfile and pick the packages in seconds. Nobody checks what they pull in. Here's how to give every agent one clean place to build from.
Ask an AI coding assistant to containerize a service. About ten seconds later you have a Dockerfile. It starts with a FROM line. It installs a few packages. It looks fine, and it probably builds on the first try.
Here's the question nobody asks: where did that base image come from, and what's inside it?
The code got faster. The inputs didn't get checked.
AI tools are normal now. In Stack Overflow's 2025 Developer Survey, 84% of developers said they use or plan to use AI tools in their work. Agents are catching up fast. In Stack Overflow's spring 2026 pulse survey, 59% said they use AI agents at work, up from 31% a year earlier.
Teams aren't careless about it. The same pulse survey found 63% rarely or never let agents run fully on autopilot. People review the code. They read the diff.
But the diff is the part you can see. The base image and everything under it come along for the ride. One FROM line can bring in a whole operating system. One pip install can bring in dozens of dependencies. None of that shows up in a pull request.
Three ways it goes wrong
of developers use or plan to use AI tools
Stack Overflow Developer Survey, 2025use AI agents at work, up from 31% a year before
Stack Overflow pulse survey, 2026of packages suggested by code models did not exist
Spracklen et al., 576,000 code samplesmalicious open source packages logged to date
Sonatype Open Source Malware Index, Q2 2026The assistant picks whatever it saw most. Models learn from public code. Public code is full of old tags, latest, and images nobody has patched in a year. The assistant doesn't know your policy. It knows what was popular.
The assistant makes things up. Researchers tested 16 code models across 576,000 code samples. 19.7% of the packages they suggested did not exist. The study found more than 205,000 unique fake package names. An attacker only has to register one of those names and wait for someone's agent to install it.
The registry is a target. Sonatype has logged more than 1.8 million malicious open source packages. In the first quarter of 2026 alone, it found 21,764 new ones. Attackers go where the pulls are. Agents pull a lot.
The code itself isn't clean either. Veracode tested more than 100 models on 80 coding tasks. The AI introduced a security flaw in 45% of cases. That's the code people do read. The base image under it rarely gets a look at all.
You can't review your way out
More reviewers won't fix this. An agent adds a dependency in the time it takes to open a ticket. If every pull needs a person, you lose the speed you bought the tool for. If no pull needs a person, you ship whatever the internet had that day.
So change where the agent gets its parts. Don't fight what it writes. Control what it builds from.
Give every agent one clean source
That's what Ghost does. Ghost is one catalog your whole software factory builds from:
- Base images, hardened and kept current.
- Language runtimes for the services you ship.
- The top 3,000 packages from the three biggest ecosystems.
- Signed, with an SBOM on every image and package.
- FIPS-enabled images for teams that need them.
Coding agents, CI runners and developer laptops all pull from the same place. Point them at Ghost, or mirror the catalog into the registry you already run, like Amazon ECR. Your clusters and pipelines pull from it like any other registry. There's nothing to install on your nodes.
For most teams the switch is one line:
# what the assistant wrote
FROM python:3.12
# what you ship
FROM ghost/python:3.12
Then tell the assistant. Most coding tools read an instructions file in the repo. Add a line like "Base images come from ghost/. Packages come from our Ghost mirror." Now the default it reaches for is the clean one.
Now you can answer the audit question
Sooner or later someone asks, "What did the agent add?" With Ghost you have an answer. Every image is signed, so you can prove where it came from. Every image has an SBOM, so you can list what's inside. It works the same whether a person wrote the Dockerfile or an agent did.
It doesn't have to cost more, either. Ghost is one flat price for the catalog, not a count of images. One Navy program moved from its old hardened image vendor to Ghost with the same coverage. Its annual cost dropped 62%.
See where you stand
If your teams ship AI-written code, start with what it's built from. We'll run a Ghost Container Economics Assessment: what you pull today, what it costs you, and what it looks like on Ghost.
Ask for an assessment and pick "A Container Economics Assessment" on the form.