Richard

Founder of RAIVEN Rational AI Ventures

[Richard @raiven_eu]

@Eric_Wallace_ @Eric_Wallace_ I applied for Daybreak blue. But there is this lingering thought that one should not need to be applying for to get special access to adversarially agent review his/her own freshly authored code (by the same coding agent).
https://t.co/oxPozsFDTL

[Richard @raiven_eu]

@sama @sama @OpenAI
I totally get the importance of getting the safety classifiers right, but could you consider allowing deep adversarial reviews of agents’s own generated code, preventing thus the rise of future security vulnerabilities in general?

[Richard @raiven_eu]

The proposal in full: the mechanism, the architecture, and what it deliberately does not solve:
https://t.co/oxPozsFDTL
The story of how we got here continues below. 2/10

[Richard @raiven_eu]

@karpathy @simonw Ahoj Andrej, pozdrav od krajana.

5500 lines in 2 hours! But who security-reviews them? If you ask an agent today to adversarially review code your own agents wrote an hour ago, the cyber classifier degrades or kills the session.

Proposal: https://t.co/oxPozsFDTL

Richard

[Richard @raiven_eu]

@simonw Follow-on effect: ask an agent to security-review code your own agents wrote an hour ago, and the post-incident classifiers degrade or kill the session. Fresh code, no target, still flagged.

Open proposal — provenance for agent-authored code:
https://t.co/oxPozsFDTL

[Richard @raiven_eu]

Feasibility question for @lorenc_dan: could generation receipts for model-authored code ride on sigstore/Rekor? Vendor's inference layer signs spans at generation; verification is cross-vendor at review time via a shared trust root. What breaks first?

https://t.co/oxPozsFDTL

[Richard @raiven_eu]

LLM vendors tightening their security classifiers has an unintended and rather ironic consequence: seeding future security bugs at a potentially unprecedented scale.
Here is why, and an open proposal on how to improve it. 1/10

[Richard @raiven_eu]

Not asking to switch the classifier off, only to improve its precision: vendors cryptographically sign the code their models generate; signatures verify across vendors at review time, so your agents' own code can be properly reviewed.
Proposal: https://t.co/oxPozsFDTL 9/10

[Richard @raiven_eu]

Open invitation, not a complaint: @karpathy @jackclarkSF @julien_c @dguido and anyone building or breaking agentic software — it gets safer when agents are allowed the engineering practices humans have relied on for decades. Adversarial self-review tops that list. 10/10

[Richard @raiven_eu]

@BitcoinArchive he speaks of bitcoin network, not 1 BTC.. in other words market cap 🧢 f the asset class

[Richard @raiven_eu]

@martypartymusic @tradingview 📈📉 need!

[Richard @raiven_eu]

@CarlBMenger @PeterSchiff 🤣

[Richard @raiven_eu]

@tradetravelchil fiat debasement. This should be taught in highschool econ

[Richard @raiven_eu]

@whaleroomtrades @kyledoops pls let me in🙃