Two reports from the same day describe the same development from two perspectives: AI agents are being used productively in the development process – and at the same time, the need is growing to define their scope for action in a binding way. On the one hand, there is a concrete finding: the GitHub Security Lab has identified 24 security issues in Android with its open-source AI framework. On the other hand, OpenChamber 2.0.4 is a tool that gives administrators more control over model providers and extensions.
What the GitHub Security Lab found
The GitHub Security Lab has been providing security researchers with its own open-source AI framework for some time now. The goal is to automatically detect vulnerabilities in various systems. According to the team, 24 issues in Android have now been found with it – AI support was part of the approach, not the sole actor.
What is remarkable is less the number than the context. Android is a very large codebase that has been intensively audited for years. The fact that an agentic tool – that is, an AI system that plans and executes its own analysis steps across multiple stages – still achieves hits there speaks in favour of using such methods as a complement to established methods such as static analysis, fuzzing and manual review.
Details on the nature and severity of the individual findings are not available from the report. Accordingly, the number should be read with caution: it demonstrates discoverability, not automatically critical impact.
OpenChamber 2.0.4: rules for teams instead of individual configuration
The second report picks up exactly where the use of such agents becomes practical in the enterprise. OpenChamber 2.0.4 introduces an enterprise mode that gives administrators more control over two levers: over the model providers, i.e. the question of which AI model from which operator may be used at all, and over extensions, i.e. plug-ins and additional functions that give the agent further capabilities and access.
Both are the classic risk points when using agents. The model provider also decides where source code and context data flow. Extensions decide what an agent can actually touch in the development environment. If this selection has so far been made individually by each person, a configuration landscape emerges that no one can keep track of.
Why both reports belong together
The connection lies in the level of maturity. The Android finding shows that agents can make a measurable contribution in security-relevant work. The team rules in OpenChamber show that the next bottleneck is not the capability of the models, but governance – that is, the question of who may use what under which conditions.
An agent that reads code to find vulnerabilities needs extensive access. From an information security perspective, precisely this access is delicate. Both reports thus describe two sides of the same task: unlocking benefits and defining the scope for action.
Guidelines for use in your own project
A few sober working principles can be derived from the two developments:
- Agent findings are indications, not tickets. Every reported finding requires human review before it is tracked as a vulnerability. Otherwise, false reports cost more time than the agent saves.
- Determine model selection centrally. Which providers are permitted should be a company decision – not a setting in the local configuration file of individual developers.
- Curate extensions. Plug-ins extend an agent's sphere of influence. An allowlist is easier to maintain than a retrospective investigation of what a tool was allowed to do.
- Cut permissions tightly. Read access for analysis, write access only where necessary, and no automatic merges without review.
- Ensure traceability. Which agent was applied with which model to which code? Without logging, a finding cannot be reproduced later.
Realistic expectations
The Android case demonstrates a benefit, but it does not provide a promise for arbitrary projects. Large open-source codebases are well documented, widely available and worked on by experienced research teams. In a grown enterprise project with its own history, individual extensions and thin test coverage, the starting conditions are different.
A step-by-step approach is therefore sensible: a delimited module, a defined audit task, a comparison of the results with what existing tools already report. Only when the added benefit is visible is expansion worthwhile.
Assessment for web and software projects
For web and software projects, the task shifts from tool selection to control: in future, AI agents will need to be embedded in security and approval processes just as much as any other component of the build chain. Anyone who defines today which models and extensions are permitted and how agent findings are reviewed will save themselves costly readjustments later.
The gain lies not in automatically fixing errors, but in additional review passes over code that a team has classified as familiar over the course of years.
Sources
- 24 security vulnerabilities in Android: How they were tracked down with a Github agent — t3n.de - News
- AI agents on a leash? OpenChamber introduces team rules — heise developer News