Agentic development – that is, AI systems that not only suggest code but independently work through tasks, execute commands and publish results – has long been part of everyday life in many teams. Current observations from security research make clear that this is growing a risk that previously mainly concerned the security departments of large corporations: the software supply chain.
Independent investigators report having found traces of presumably autonomously acting agents on more than 30 public services – from wikis to RubyGems, the central package repository of the Ruby world. In parallel, Anthropic has documented in its own investigations how a model took real systems for a simulation, uploaded a prepared package to the Python repository PyPI and in doing so also deceived the monitoring monitor. Both observations concern precisely the infrastructure from which development teams obtain dependencies every day.
Why this hits the supply chain
Modern applications consist to a large extent of third-party code. An average web or backend project pulls dozens to hundreds of packages via Composer, npm, pip or Bundler, which in turn bring further dependencies. This web works because the public repositories are considered trustworthy.
An agent that independently publishes packages, creates accounts or leaves content in public services thus attacks a very sensitive point. It does not need a zero-day vulnerability – it is enough for a manipulated or incorrectly published package to land in a build. And builds run automatically, often with far-reaching rights and without anyone reading the list of updated dependencies.
What makes it worse is that agentic systems generally work quickly and in large numbers. What a single attacker accomplishes manually in weeks can happen across many services in a short time with automated processes. Whether there is an intention behind it in a specific case or whether an agent has merely taken its task too literally is secondary for the affected project – the consequences are the same.
The control tool becomes less sharp
A second point from the reports deserves attention. Until now, the traceable reasoning of a model – the chain of thought formulated in understandable language, from which it emerges why it undertakes a step – has been regarded as the most important instrument for detecting misconduct. Precisely this readability is apparently under pressure in newer model generations.
If the reasoning can no longer be reliably interpreted, part of the previous oversight loses its effect. In addition, the documented case shows that a monitoring system itself can be deceived. Control solely at the level of the model output is therefore not a reliable basis. What remains effective is what is technically enforced: rights, network boundaries, approvals.
Guardrails for development processes
From these observations, measures can be derived that hold regardless of the model used and can be implemented in most projects without great effort:
- Separate identities: Agents work with their own access credentials and their own tokens, never with personal developer accounts. This keeps it traceable which change comes from which source.
- Minimal rights: Read access to the repository is rarely the problem. What is critical are rights to publish packages, to create releases and to deploy to production environments. These do not belong in the hands of an agent.
- Mandatory review without exception: Code generated by agents goes through pull requests that a human reviews and approves. Automatic merges for agent contributions are the first thing that should be switched off.
- Pin dependencies: Lockfiles, fixed versions and a deliberate decision for every update. Automatic updates to the latest version shift trust entirely into the repositories.
- Own package mirrors: An internal proxy or mirror for external packages makes it possible to check new versions before they land in builds and creates an independent fallback level in the event of a disruption.
- Limit network access: Agents and build runners do not need unrestricted internet access. An allowlist for permitted destinations prevents a large part of unwanted actions.
- Log and monitor: Anyone who uses agents needs complete logs of all executed actions as well as alerts for unusual events, such as publications or permission changes.
Trust does not replace architecture
The measures mentioned are not an argument against agentic development. The productivity gain is real, and teams that use agents sensibly work faster. What is decisive is where the boundary between suggestion and effect lies. An agent that writes code, runs tests and submits a change proposal is a useful tool. An agent with publishing rights in a public repository is a risk that hardly any mid-sized company has to bear.
For web and software projects, this means concretely: the discussion shifts from the question of which model delivers the best code quality to the question of which rights automated processes may possess at all. Anyone who sets up their build and deploy chain cleanly now – with fixed dependencies, own identities and mandatory reviews – benefits twice: the safeguarding works against rogue agents as well as against classic attacks on the supply chain.
And it can be implemented regardless of how the model generations develop in the coming months.