Principal @ Security Simulations. Policy and research at Princeton and CSU. Former CISO: LinkedIn, Google Devices & Services, Confluent. Ex @stake / Matasano.
Joined July 2008
- Tweets1.9K
- Following716
- Followers1.4K
- Likes865
Here are the slides from my AI Security & Privacy guest lecture for CSU’s Practical AI for Lawyers class. The slides include interactive demos, so click around!
coryscott.github.io/csu-prac…
Great read on where OpenAI is at capability-wise on isolation primitives for sandboxing training environments. Good steps forward, including accepting tradeoff of no Internet access. Would like to see more effort in perimeter service security that isn't just more red-teaming.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
When bad CAs give you lemons, make lemonade... a quick patch to mitmproxy and now all communications from the IoT devices on my test bench without cert pinning are easily visible without needing to install my own root cert. Thanks eMudhra!
Besides the CVP update, Anthropic also released updated best practices on agent containment, which includes sandboxing, agent monitoring, and a sandbox escape classifier if you're rolling your own harness via the API path. Useful even if you're not in CVP. support.claude.com/en/articl…
Cory Scott retweeted
Meet Rob Stucke, an Information Security Engineer on our Detection Response Engineering Team! 👋
We sat down with Rob to hear all about his day-to-day as well as what he and his team are building.
Want to get involved? Check out our openings!
👉 cnfl.io/444uJB8
"Cybersecurity professionals cannot collectively sustain a model where most of their resources are spent reacting to emerging problems like a steady stream of vulnerabilities." A sobering stare-into-the-abyss moment in the CSRB log4j report. cisa.gov/sites/default/files…
ALT Screenshot from the CSRB report
The state of Bug Bounties in 2022 in one tweet.
Reddit disclosed a bug submitted by dingleberryfarts: hackerone.com/reports/153815… #hackerone #bugbounty
How many concurrent Google Calendar chimes does it take before you realize you need to close some tabs? My office sounds like an arcade at 10 minutes before every hour.
After the third Smiths track within a single hour during our Valentine's dinner out, I had to give it up for the mad lad who crafted the playlist. Well played.
A reminder to all security and privacy professionals about our obligation and duty to stand up and “shout from the cheap seats”.
His work on computer crime research was way ahead of its time. He wrote a book on Computer Crime in 1976 (!). ojp.gov/ncjrs/virtual-librar…
A critical point. Apple is building a bulwark to resist more intrusive and disruptive regulation. They were on incredibly shaky ground before this announcement. They've likely bought themselves years of negotiation time with this move.
This quoted post is unavailable.