Risky Business Features

Risky Business Features

By Risky Business MediaTechnology
Download on the App Store

All Risky Business Features episodes

  • How private LLM inference actually works

    In this podcast episode James Wilson chats with Tinfoil co-founder Tanya Verma about how you can run a powerful LLM in the cloud without the inference provider seeing your prompts.

    Tanya talks James through how private inference works, from trusted execution environments and hardware attestation, to TLS termination and GPU isolation. Customers can verify the exact code and model processing their data, while Tinfoil and its infrastructure providers remain locked out.

    That’s clever engineering… but who really needs it? Is private inference only useful if you’re doing something bad, or will it become a privacy baseline like TLS? James and Tanya discuss the costs and trade-offs, and how open weights make private inference more transparent and trustworthy.

    Show notes
      1 hr 24 min
    • Benchmarks, borders and the true cost of AI regulation

      The US government is flirting with the idea of regulating most open weight models out of existence. What would that mean for everyone who’s currently using them?

      Policymakers who just look at the benchmarks and token costs might think frontier and open weight AI models are all interchangeable. That they can just banhammer them without causing any disruption. That’s far from the truth.

      In this solo podcast, James Wilson looks beyond the US vs China AI race rhetoric and instead games out what the real world consequences of the US government slapping bans on AI models could be.

      Show notes
        40 min
      • Fortibleed: The bleeding edge of AI cybercrime

        In this podcast episode SOCRadar CISO Ensar Seker and James Wilson chat about the company’s deep dive into the Fortibleed campaign. A small investigation into a curiously open directory on an unknown server expanded into the discovery of an attack that targeted 400,000 Fortinet devices.

        As Ensar says, each time the SOCRadar team pulled a single thread, it led to a tapestry of AI-enabled cybercrime. They uncovered custom initial access, persistence and packet sniffing tools, as well as direct links to the INC and Lynx ransomware operations. Most interesting though is the use of AI to design, implement and operate all aspects of the campaign across a team of 20 individual actors. Operating more like a modern software company than a traditional cybercrime gang, Fortibleed serves as our first in-depth look at the future of cybercrime.

        Show notes
        • SOCRadar Links FortiBleed Campaign to INC and Lynx Ransomware Operations
        • 49 min
        • What to do 'til the bugpocalypse gets here

          In this podcast episode Brad Arkin joins James Wilson to discuss how defenders can get ahead of the late-running bugpocalypse. While we’re confident the offensive cybersecurity capabilities of frontier and open-weight LLMs are real, attackers don’t yet seem able to fully utilise them. This creates a window of opportunity for defenders to tackle the threat.

          There are a few well-funded and seemingly overlapping industry efforts under way including Athena, Akrites, and Patch the Planet. But, as Brad says in this interview, there’s too much focus on fixing bugs and traditional vulnerability triage, and not enough on exploring how to make entire classes of vulnerabilities inert.

          Show notes
            45 min
          • Mythos on your desk? Using local LLMs for code reviews

            In this podcast episode James Wilson chats with Karsten Nohl about his research into using local LLMs to replace cloud AI in security code reviews.

            In essence, Karsten created a hybrid code reviewing system where both cloud and local models are used to orchestrate, triage outputs, and write reports. In this system, only the local LLMs have source code access, with the cloud models used to manage the local models.

            In this “source-local” review technique, the source code never leaves the local endpoint, which is a requirement for some reviews. But funnily enough, Karsten was able to use this system to generate findings that were as impressive as when using frontier models directly.

            In a nutshell, Karsten proved it’s possible to use locally-hosted, open-weight models running on commodity hardware to produce findings comparable to those discovered by frontier cloud models.

            This episode is also available on YouTube.

            Show notes
            • Beyond Fable: Can a Local LLM Replace Cloud AI for Security Code Reviews
            • Mythos smythos! How to find 0day with lesser models
            • 1 hr 12 min
            • Pitching security startups to VCs in the AI era

              In this podcast Patrick Gray and James Wilson chat with Decibel Partners founder and Managing Partner Jon Sakoda to talk about pitching cybersecurity startups to VC firms in the AI age.

              Coding agents and large language models have made it easier than ever to create software products, but despite this, the bar for what interests an investor is still largely the same. Everyone can run the marathon, but it’s usually the same few folks who finish first.

              So tune in to hear Jon share with us his wisdom on when to start the conversation with investors, how to leverage the experience of the founder community, and what founders should watch out for.

              This episode is also available on YouTube

              Show notes
                36 min
              • How using open weight models can blow up in your face

                In this podcast episode James Wilson and Brad Arkin talk about how to safely use open weight large language models in the enterprise. The cost of frontier models was already driving interest in freely available open weight models like DeepSeek, Kimi and Qwen. But now the US government is forcing Anthropic to pull its Fable and Mythors models from the market, the argument for having greater control over your own AI stack is stronger than ever.

                But as you’ll hear in this episode, the model itself is just one component of the complex tech stack you’ll need to spin up if you want local inference. There’s a lot of moving parts, each of which comes with its own supply chain risks.

                So whether you’re hosting these models on your own hardware or via a SaaS provider, there’s a lot to ponder!

                Show notes
                  44 min
                • The state of the art in AI model jailbreaks

                  In this solo podcast episode, James Wilson breaks down the current state of AI model jailbreaks.

                  If you’ve somehow missed the story, last week Anthropic released its Fable 5 and Mythos 5 models to the public. In the name of safety, both models were guardrailed up the wazoo, but that didn’t stop a bunch of jailbreakers from figuring out how to bypass at least some of their safety restrictions.

                  In response to these guardrail bypasses the White House issued an export control directive on the models, citing national security concerns. But was the Trump administration right to do this? Do these jailbreaks represent a threat to the security of the USA, or was the export restriction overkill? Tune in to find out!

                  Show notes
                  • Pliny the Elder on Fable 5 Jailbreak
                  • whoJumper's response to Pliny
                  • ConfusedPilot: Confused Deputy Risks in RAG-based LLMs
                  • 53 min
                  • Why NPM v12 won’t stop supply chain attacks

                    In this podcast episode, James Wilson is joined by Open Source Malware Security co-founder Paul McCarty to talk about the supply chain attack mitigations coming in NPM v12.

                    NPM disabling (by default) auto-run install scripts and dynamic dependencies is a positive step forward… but it’ll take years for this new version to be adopted, and these changes do nothing to prevent malicious packages being imported into projects.

                    Further, Paul thinks disabling these features by default will introduce friction that will cause them to be re-enabled. When the choice is “this builds” and “this is less prone to malware”, the former will always win.

                    Show notes
                      39 min
                    • Everything is getting much worse, much faster

                      In this podcast Brad Arkin joins James Wilson to talk about how the fear of being left behind in the AI era means enterprises are taking risks that would have been considered insane just a couple of years ago.

                      Fears around outages or being hacked have been trumped by fears of being labelled an AI laggard.

                      So where are we all going? Say hello to tech debt-riddled, vibe-coded apps, crazy dependencies on AI providers, and an emerging threat landscape that can’t be mitigated by a contemporary SOC. Sounds like fun, eh?

                      Show notes
                        24 min