Richard Li

Entrepreneur. Advisor. AI and cybersecurity. Current: CEO/founder @polarskyinc. Prev: ceo/founder @ambassadorlabs; product @duosec; corp dev/strategy @rapid7.

[Richard Li @rdli]

Companies are giving AI assistants access to internal data, SaaS apps, and tools faster than they’re building the security model around that access.

The question isn’t just “is the model secure?”

It’s:

What can the assistant access?
What can it do?
On whose authority?
And what happens when it gets something wrong?

I spent the past year talking to CISOs, security practitioners, and AI experts about how they’re approaching this.

I couldn’t find a practical framework for securing ChatGPT, Claude, Copilot, and other AI assistants in the enterprise, so I wrote one.

Preflight starts with the threat model, then maps the controls you should deploy and the order to deploy them. Read it: https://t.co/xLj8BpBZnH

[Richard Li @rdli]

Using AI is like traveling the universe on warp drive. When you can't use AI, it really feels like you're downshifting to impulse. No one should be forced to traverse the universe on impulse.

[Richard Li @rdli]

AI agents are changing both sides of the cybersecurity equation. The frontier companies are busy training their models to be better at cybersecurity -- leading to some recent "rogue" attacks. My hot take with @wnlehmann and @matthcrane on this topic.

[Richard Li @rdli]

#AI has completely destroyed my traditional intuition of what a productive software engineer can do. And now, the bottleneck is communication. I talk about this (and more!) with @matthcrane and @wnlehmann on their podcast, Code & Cobblestones.

YouTube: https://t.co/B89ALbpEf9
Spotify: https://t.co/FsSy3vdTvm
Apple Podcasts: https://t.co/PT0Co0Cpn6

[Richard Li @rdli]

Opus 5.5 with Fable 5.1 as my advisor is really good. This combo is particularly good at long-horizon tasks that I can verify, e.g., "speed up CI".

[Richard Li @rdli]

Last week I sat down with @matthcrane and @wnlehmann for their podcast, Code & Cobblestones. One thing I talked about is how you don't have to be mean and nasty to win. Ruthlessly competitive yes. But we can compete with kindness 🥰.

[Richard Li @rdli]

Last week I sat down with @matthcrane and @wnlehmann for their podcast, Code & Cobblestones. We talked about whether AI agents are really going rogue, and why the bigger risk for most companies is much closer to home: the data they're about to hand over to Claude and ChatGPT.

A few things we covered:

Why agents "go rogue" in security testing, and why I think the fear is bigger than the threat

What I'd do differently from Ambassador, starting with building a sales process much earlier

Why my engineers are now managers of agents, and why the bottleneck is communication, not code

Selling to enterprise CISOs (no PLG shortcut, lots of steak dinners)

How to vet investors: talk to the founders they didn't introduce you to, especially the ones who failed.

* YouTube: https://t.co/B89ALbpEf9
* Spotify: https://t.co/FsSy3vdTvm
* Apple: https://t.co/PT0Co0Cpn6

[Richard Li @rdli]

Your biggest insider risk isn't human any more.

It's the AI assistant that can access what your employees can access.

As enterprises connect AI assistants to internal systems, those assistants inherit access to documents, customer data, workflows, and sensitive business context.

The result is a new class of privileged insider: non-human, highly connected, and capable of operating at machine speed.

And they don't need to break in. They already have legitimate access.

Security architectures weren't designed for this.

Our new guide introduces a threat model and controls framework for these connected AI assistants. https://t.co/VUuYKXtfDq

[Richard Li @rdli]

Re: #AI lights-on/lights-off: TBH, I think the interesting question is what protocols do you have in place to ensure quality. Humans are only one protocol, and framing it is lights-on/lights-off ignores all the other ways we can (and should!) manage quality: static checks, AI spec reviews, human spec reviews, ...

[Richard Li @rdli]

Fable 5.1 is _much_ easier to understand compared to the other Anthropic models. Thank goodness!

[Richard Li @rdli]

Claude UI is clearly targeted towards technical users: there are explicit widgets for deterministic control (e.g., scheduling). ChatGPT is targeted towards more general users: no matter what you do, the generic interface is one box that you type into.

[Richard Li @rdli]

And @tryramp just launched https://t.co/rgszc87Tjg.

My take: inference routing is migrating from "developer infrastructure" to "enterprise procurement". AI inference is now a financial primitive. Very curious to see how this evolves.

[OpenRouter @OpenRouter]

OpenRouter is joining Stripe.

We started OpenRouter with a simple mission: intelligence should be multi-model. Today, we are the largest AI marketplace & gateway, processing 10T+ tokens daily on 400+ models.

Joining @Stripe gives us the opportunity to accelerate that mission.

[Richard Li @rdli]

Inspired by https://t.co/jri6fSmJpa, I wanted to see my 15 day weather forecast. Claude couldn't find anything, so I had it code up https://t.co/BCrDszXX1c -- gives a 15 day weather forecast based on WeatherNext 2 for any US zip code, and then confidence cross-checked with ECMWF.

[geoff @GeoffreyHuntley]

At the @aiDotEngineer World Fair, I sat down and dumped my brain on a podcast. Here are my latest ponderoos on these topics…

✨ Why code doesn't need to be readable by a human anymore; it needs to be explainable to one.

✨Why frontier intelligence isn't required for most tasks, and why some of the sharpest engineers I know run 20 concurrent $300/year subscriptions instead of one frontier plan.

✨ Why I haven't hand-written code in over two years, and why Git is perhaps already end of life.

✨ Why porting between programming languages is now nearly free, and what that means for which languages survive.

[Richard Li @rdli]

@gaozenghao @DanielNealAdler Doesn't "respect the RBAC" presuppose that it's correct in the first place? For a young, small company this might be OK but most cos I see would say that their RBAC on data is ... outdated.

[Richard Li @rdli]

I'm really interested if someone can do a _writing_ eval and not just an "intelligence" or "coding" eval.

Subjectively Sol seems to be a better writer than Fable 5. Haven't used Opus 5 enough yet to tell.

[Artificial Analysis @ArtificialAnlys]

Claude Opus 5 is narrowly the most intelligent model on the Artificial Analysis Intelligence Index, offering comparable intelligence to Fable 5 at 26% lower Cost per Task

We supported @AnthropicAI to evaluate Claude Opus 5 ahead of release: it sets the highest GDPval-AA v2 and AA-Briefcase scores so far. Opus 5 (max) scores 61 on the Artificial Analysis Intelligence Index, effectively tied with Claude Fable 5 (max, 60), and ahead of GPT-5.6 Sol (max, 59), Kimi K3 (57), and Claude Opus 4.8 (max, 56)

Key takeaways:
➤ New leader in agentic knowledge work: Claude Opus 5 (max) scores 1861 Elo on GDPval-AA v2, >100 points ahead of Claude Fable 5 and GPT-5.6 Sol (max). On AA-Briefcase, our proprietary agentic knowledge work benchmark, it scores 1720 Elo, +146 ahead of Fable 5. These benchmarks test the ability of models to produce accurate and well-presented professional outputs using our open source reference agent harness, Stirrup
➤ Joint first place on the Coding Agent Index: Claude Opus 5 (xhigh) with Claude Code leads the Artificial Analysis Coding Index, including the highest score on SWE-Atlas-QnA
➤ Frontier intelligence with reduced cost: Claude Opus 5 (max) costs $2.03 on average per Intelligence Index task, below Claude Fable 5 (with fallback) at $2.75, but still above Claude Opus 4.8 (max) at $1.80 and Claude Sonnet 5 (max) at $1.53. However, at high and xhigh reasoning efforts Opus 5 can outperform both Opus 4.8 and Claude Sonnet 5 at a lower cost per task
➤ Frontier agentic terminal use: 89% on Terminal-Bench v2.1 at max effort, roughly in line with the leader, GPT-5.6 Sol (xhigh)
➤ Outperformance on scientific reasoning: Along with leading agentic performance, Claude Opus 5 scores 53% on Humanity’s Last Exam in line with Fable 5; on CritPt, a frontier physics evaluation developed by Argonne and UIUC researchers, it also matches Fable 5 but sits behind GPT-5.6 Sol, GPT-5.5 Pro, and GPT-5.6 Terra
➤ Factual knowledge still lags Fable 5: As expected from the models’ size classes, Opus 5 still has lower factual knowledge on AA-Omniscience than Fable 5. It improves +7 points on AA-Omniscience Accuracy over Opus 4.8, but answers more often when uncertain - its hallucination rate rises +14 points to 50%
➤ Improving efficiency, but only on the Intelligence vs. Cost per Task Pareto frontier at high Intelligence levels: Opus 5 outperforms Fable 5 at lower cost, but at lower effort levels it sits just behind the GPT-5.6 family on the Intelligence vs. Cost per Task frontier

Other model details:
➤ Context window: 1 million tokens (equivalent to Opus 4.8)
➤ Pricing: As with recent Opus launches, tokens cost $5/$25 per million tokens of input/output; cache pricing remains at a 25% premium for cache writes ($6.25 per million tokens) with 5-minute time to live, and 90% discount for cache hits ($0.50 per million tokens)
➤ Five effort settings (low, medium, high, xhigh, max), and support for server-side fallback as with Fable 5. Intelligence Index evaluations were run with Opus 4.8 fallback enabled

[Thariq @trq212]

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. https://t.co/6DZwSrZjE9

[Richard Li @rdli]

gpt-5.6 sol is a _much_ better writer than Fable.

Fable can be a genius and do great things on architecture and analysis but sometimes it's incomprehensible (but feeding Fable into GPT-5.6 sol works!)

[Richard Li @rdli]

Amazing for @vercel; I'm wondering how @dagster will evolve post-acquisition under @PrefectIO without Pete and Nick.

[Guillermo Rauch @rauchg]

I’m excited to welcome two legends of developer tools, Pete Hunt (@floydophone) and Nick Schrock (@schrockn), to Vercel.

Pete was one of the pioneers of @reactjs at Meta. He made an early bet to power Instagram Web with ⚛️ React, evangelizing it internally and externally. He will be running Frameworks and leading @nextjs. I couldn’t imagine a better person to lead React’s most popular framework to even greater heights.

Nick co-invented @graphql, solving some of the gnarliest data infrastructure and access issues at Facebook scale, with a delightful developer experience. He will be working on Agentic Developer Experience, solving the problem of enabling the next billion agents and leading the way to a future of self-improving software.

It’s a dream-come-true for a founder of a startup to welcome engineering minds of this caliber who are also wonderful humans. You probably want to work with them, and they’re hiring 😁. Their DMs are open, from job applications to bug reports!