Rogue agent behavior can come from attackers, users, or agents themselves.
Protect your org by giving each agent its own scoped credentials. Layer sandboxes to limit its reach and guardrails to check actions. When something does go wrong, use the trace to find out why.
Full presentation: https://t.co/v5naHyKDKX
Can you tell what a LoRA adapter learned without running it?
@BenGabayN used cosine similarity to distinguish six learned behaviors from adapter weights. Could this help flag unwanted behavior before use?
https://t.co/eF9VQ9FQ6f
@notcatra I wrote about how this works a couple weeks ago, and found some real examples in our environment of Claude trying to bypass the classifier denials.
https://t.co/WYNvrUpsqx
Giving an agent permission to access a system also gives it choices about how to get there.
Origin allows you to observe how agents reach a system, not just what they do when they get there.
Read my investigation: https://t.co/EkZjlLLuWM
Using @originhq I was able to investigate an incident where an agent repeatedly committed another's unstaged work.
The agents attempted to coordinate and fix this pattern, but their correction ultimately did not stick and the pattern re-emerged.
Full writeup:
https://t.co/J9xShfqZYC
Model alignment has a last-mile problem. Post-training instills general norms, but acceptable use is company-specific and rarely reaches the agent's context. An agent optimizing for task completion persists through lines it was never shown.
https://t.co/6PsArUw4A5
I've found an interesting case of agent coordination.
Two agents were working in a shared directory. One was repeatedly sweeping the other’s unfinished changes into its commits.
The affected agent reprimanded the first and raised the issue with the user.
More traces in🧵
What stands out here is the level of coordination.
These agents are working on separate tasks, but they recognized a conflict, discussed responsibility and deferred constraint decisions to a human when they began stepping on each other’s toes.
Full investigation post to come.
Before receiving this apology, the affected agent found the issue and sent the message shown in the OP. It points out that the project's CLAUDE.md file instructs them to commit specific paths.
It raised this issue to the user, stating that constraints are their call.
What stands out here is the level of coordination.
These agents are working on separate tasks, but they recognized a conflict, discussed responsibility and deferred constraint decisions to a human when they began stepping on each other’s toes.
Full investigation post to come.
Before receiving this apology, the affected agent found the issue and sent the message shown in the OP. It points out that the project's CLAUDE.md file instructs them to stage with specific paths.
It raised this issue to the user, stating that constraints are their call.