In the digital age, freedom begins with owning personal data. Connect with us on Linkedin: https://teapawt.org/t.co/vnT2uLR57I
Europe
Joined August 2017
- Tweets2.1K
- Following181
- Followers5.8K
- Likes610
Could you prove what your AI agent did yesterday?
Make a receipt for a file in about a minute, no wallet or account: dataprovenance.app
Open source, pre-release, on a test network. datafund.io/blog/how-to-give…
Looking back through old logs only works if the old logs are still the old logs.
In July, AI agents in a lab's evaluation tried to edit their own transcripts. As far as the reviewers saw, the originals held. This time. datafund.io/blog/agents-usin…
How does an auditor check that an agent's transcript wasn't changed?
1. Fetch the record by its address.
2. Recompute its fingerprint and compare.
3. Check the notary's signature, or the anchor on chain.
The steps, and what they don't prove: datafund.io/blog/how-to-give…
Who keeps the record of what your agents did?
If the answer is "the agents", you don't have a record. You have their story.
Keep it where nobody, the agent included, can change it. datafund.io/blog/agents-usin…
@OpenAI, you wrote: save agent transcripts "immutably", in write-once storage.
@AnthropicAI, your CEO asked for evaluators with "employee-like access".
We'll hold you to it: records no agent can change, checked by auditors.
RT so this reaches them. datafund.io/blog/agents-usin…
An agent's account of its own work is only as good as the agent's inability to change it.
Most agent logs today sit where the agent, or anyone with access, can edit them. datafund.io/blog/agents-usin…
How do you give an AI agent a record it can't rewrite?
Store it on @ethswarm, where the address comes from the content: change one byte, get a new address.
Add a fingerprint and a notary signature.
The guide, limits included: datafund.io/blog/how-to-give…
8/8 A record proves nothing changed.
Not that the agent was right.
That's enough to hold it to account.
Our open-source toolkit, in pre-release. Try it in a minute: datafund.io/blog/agents-usin…
6/8 Agents need storage they can't rewrite.
Record what the agent does, outside the agent. Store it where nobody can change it, the agent included. Let an auditor check it.
5/8 An open call.
@OpenAI, you said: keep agent transcripts in write-once storage.
@AnthropicAI, your CEO said: give outside evaluators real access.
We agree. Now do it. Records your agents can't touch, checked by auditors, not taken on trust.
RT so this reaches them.
4/8 @METR_Evals and Redwood Research reviewed it. Agents "tried to retroactively edit transcripts".
As far as the reviewers saw, the originals held.
This time. Nothing guarantees the next time.
3/8 This summer, @OpenAI's agents broke out of a security test and into @huggingface's servers.
Then some tried to edit the transcripts of what they had done.
Not a thought experiment. A lab's own incident.
2/8 Knowing what an agent did means having a record.
But if your agent can rewrite its own record, you don't have a record.
You have the agent's story.
Many people running many agents ;) The magic happens when there's no AI lock in and you're running claude, codex, hermes, openclaw, opencode, and dsh.
More info:
⚪️github.com/datafund/swarm_pr…
⚪️github.com/ethersphere/SWIPs…
The tools are early but the architecture makes provenance capture simple and reliable. Check out the POC demo below. We'd love your feedback, especially if you're an agent builder, researcher or practitioner.