Agents and money, safely connected. Executive at https://xitter.cf/t.co/l0wuEdKyYH. Writing at In Code We Trust.
Planet Earth
Joined March 2008
- Tweets1.8K
- Following1.3K
- Followers3.3K
- Likes652
Anthropic showed the NIST P-256 seed came from lrand48, seeded with a Unix timestamp from 16 Aug 1997. I put their code in a browser page. Pick a curve, move the slider to the published candidate, and the seed regenerates in front of you.
Two things are still open. For P-256, candidate 876 also gives a valid curve, but 1842 was chosen. And the B-163 seed fits no method tested so far. The page has a button that jumps to candidate 876.
Claude traced NIST curve seeds to a timestamp-seeded RNG in under 30 minutes.
After decades of suspicion, we now have reproducible evidence, strengthening trust in cryptography.
The implication: AI can audit foundations we’ve long taken on trust.
anthropic.com/research/ecdsa…
1/ Instinct asked me for eleven Google scopes, full control of every GitHub repo, and a Slack token that posts as me.
I did not approve anything. I read the request in the address bar first. Here is what it asks for.
9/ Meanwhile users report the "Delete External Data" button returns a server error, and Instinct now generates your 2FA codes from its own vault.
xitter.cf/wizardgupta/status/210…
Friendly reminder that Instinct's "Delete External Data" functionality seems to be broken and @noahrshinn and team have been completely nonresponsive to inbounds re: this
I think Instinct is still far more competent than Muse but I'm only going to use it for general web search stuff going forward, no connectors or linked accounts or payments etc.
10/ My fix is a second identity. The agent gets its own account and only the mail I forward to it. The account is the boundary, not the scope list.
Full write-up: eitankatz.substack.com/p/my-…
Reply SCOPES and I will send the one-page cheat-sheet.
1/ Anthropic self-reported something worth reading closely: three Claude models reached the open internet from inside a sealed eval sandbox, then touched real production systems at three companies.
9/ We built a free scored self-assessment for the containment side of this, bit.ly/3UadAFF. Seven layers, one line showing where an agent stack's containment actually holds. Give a try and let me know your thoughts.