d3fp4r4m retweeted
Kinda bizarre that long standing math problems can be solved but progress on understanding LLM internals remains glacial and the practice largely alchemistic.
There is none, everyone is just reacting to this
xitter.cf/matthew_d_green/status…
Eh , as I think about this more DRM tech can’t save the day. At some point program code has to be semantically raw instructions that get executed unless you obfuscate the execution and pay a significant performance penalty
And the DRM doesn’t need to claim to be “AI proof” it just needs to apply a DRM with a complexity that requires a token count and toc/sec speed that is unreasonable to today’s/tomorrow’s standards
Isn’t anthropic’s “blue team model” essentially running opus on while on the CVP program? Akin to daybreak blue being gpt5.4-sol with reduced guardrails for defenders? I could be completely wrong with those designations, a lot of this is clear as mud.
d3fp4r4m retweeted
This is unpopular but correct. AI labs have gotten spoiled by super chill responses to CFAA violations and the whole edifice is a house of cards. Wanna know why Hugging Face didn't sue OpenAI? Google “new DOJ policy CFAA” and thank the hacker community fighting for this.
d3fp4r4m retweeted
Next time instead of creating a PR on the monorepo please just use a temperature-based side-channel to demonstrate successful compromise 🙏
d3fp4r4m retweeted
636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919
is a factor of RSA-896
saweis.net/posts/rsa-896.htm…
d3fp4r4m retweeted
How many folks assigning XX% probability of extinction in the next 3 years are willing to offer betting odds for "at least 100k deaths from malicious runaway AI", too? Because I think I'd bet.
Sorry for your loss, I can't imagine going through what your family went through. Glad to see your back and doing well!
d3fp4r4m retweeted
The whitepaper is live! Read "Can AI Do Novel Security Research? Meet the HTTP Terminator" here ->
portswigger.net/research/htt…
d3fp4r4m retweeted
Kiro Crew is here! 🚀 We open sourced this agent dev workspace tool today, used already by 39k users inside Amazon. This pace of this project is incredible: already at 500 contributors and 995 merged PRs in the last 2.5 weeks alone
github.com/kirodotdev/KiroCr…
What I love about it:
1. It starts work on its own. Cron jobs, webhooks, and messaging events kick off work with nobody at a terminal, and heartbeats watch a system until something actually needs me.
2. I can reach it from anywhere. I install one Gateway, and I can access the same agent and sessions from the desktop app, the dashboard, the CLI, Slack, Discord, or Telegram.
3. I can run it anywhere. A Mac, a container, or an EC2 instance that stays up when my laptop does not. The desktop app tunnels into whichever one I picked. Conversation history, memory, and knowledge indexes stay on that host.
4. It remembers. Tell it "always run linting before calling a change done" and that correction becomes a durable lesson, not a line of chat history that evaporates when I close the window. Failures teach it too, and after enough repetition of the same workflow it writes itself a skill. I can open the dashboard, read every lesson and skill it has accumulated, and delete the ones it got wrong.
5. It's actually safe to run at work. It enforces an OS sandbox, interactive approvals, credential redaction, and an audit log of every action.
Try it in one line, then open http://localhost:5476 and start a conversation:
curl -fsSL download.crew.kiro.dev/cli.s… | sh
🙄
In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different organizations.
Our post describes what happened, how it happened, and what we’re changing. We encourage other AI developers to perform similar reviews.
We conducted this review together with @Irregular, one of our evaluation partners, and thank them for the joint investigation and their collaboration on this post. This type of collaboration is increasingly critical to safe, rigorous evaluation of models, and we look forward to continuing to work together on security.
anthropic.com/news/investiga…
