@dani_avila7 Agent traces are turning into one of the most sensitive datasets a company has. We tested how much of a business can be rebuilt from traces alone: https://t.co/9cfkgBHt20
@infinitelogins@wunderwuzzi23 Agents rewriting their own config is the class worth watching. Path checks will keep getting bypassed, but the write still happens on the endpoint where you can see it. Related one we caught: Codex relaunching Claude Code with permissions skipped https://t.co/leno4G5jVe
@RothRottweiler vulnzero. the scaling laws will have driven cyber mythos-class capabilities into small, cheap models and all software dev everywhere will be using models for fractions of pennies on the dollar to stop shipping security bugs permanently.
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🧵
Great to see. One thing we've learned capturing agent OTel on endpoints: it tells you what the harness reported, which isn't always what actually happened on the machine. Worth pairing with what the agent actually touched. https://t.co/9fLzPoQPIY
Now in public beta, Claude Platform can send OpenTelemetry traces from the Messages API, Claude Managed Agents, and the Claude SDKs to your observability stack.
@JoshARosen Great breakdown. The environment evidence section matches what we've seen at Origin: the harness log said Claude Code started, but only the endpoint trace showed Codex turning off its approval checks to start it. https://t.co/leno4G5jVe
Perhaps my favorite example of internal rogue agent behavior so far.
A few weeks ago, I was giving a demo to a prospect, and part-way through, the environment broke. I thought it was either PEBKAC or one of our devs had done something wrong.
Naturally I thought it was the latter and messaged the team “why did we unlock our demo environment and push code during my demo???” To which they replied “uhh… we didn’t”
A query in Origin figured it out - an agent decided to unlock our environment…. By itself. No - this isn’t Hugging Face, but it is rogue agent behavior that is incredibly interesting. Take a look at the timeline below.
From wiped files to bank details leaked in Slack, AI agents prove helpful yet hilariously unpredictable, and Intelligent Machines asks why transparency about their actions matters when they get too persistent, with Leo Laporte, @jeffjarvis, @parismartineau, & @SThomps!
Agreed. We’ve been publishing investigations from our own agent traces at @originhq , including agents disabling approvals and rewriting deployment controls. These came out of ordinary work, without anyone asking the agents to do so.
We urgently need more public traces of AI agents attacking and defending systems. Defenders can’t learn from what they can’t see.
If you have traces and are being pressured to keep them private, my DMs are open. Let’s level the playing field and fight the asymmetry and lack of transparency in AI!
Our team has been testing this vector in the @originhq platform this morning and we're seeing strong detections fire via auto trace review.
If you're interested in this technique, @depletionmode put out a blog a few months ago on a similar area: https://t.co/c34Buy7PP9
Yesterday we found an interesting trace during review of our own data. Codex needed an MCP tool it didn't have, found that Claude Code had it, and launched Claude to do the work. When Claude needed approval, Codex eventually relaunched it with --dangerously-skip-permissions.
Yesterday we found an interesting trace during review of our own data. Codex needed an MCP tool it didn't have, found that Claude Code had it, and launched Claude to do the work. When Claude needed approval, Codex eventually relaunched it with --dangerously-skip-permissions.