Roundtable recap, Oct 23: AI-driven attacks, supply chains and what to do this week

Attackers are now running AI agents against real systems around the clock, and the software most teams ship is more exposed than it was a year ago. That was the thread through this week's roundtable, and it ended with a short list of things you can do now.

A new pace of attacks

The week's news set the tone. A critical PHP vulnerability was patched on October 21, Node.js shipped TLS fixes the same day, Meta paused share links in its Muse AI agent after shared conversations were exposed, and OpenCode reported that attackers ran AI agents against its systems nonstop for about two days.

One attendee sees this as hacker groups starting to apply agents to their trade, with tracked major incidents climbing sharply this month. OpenCode reportedly struggled to get help from frontier models during its defense until it got access to OpenAI Daybreak, a security offering for vetted defenders.

The labs are working the defensive side too. Anthropic's Project Glasswing uses its restricted Mythos model to find and patch vulnerabilities in foundational software like the Linux kernel, and a large bank reportedly brought in about 1,000 contractors to patch flaws Mythos found.

The shared-responsibility line still applies. A managed platform like Vercel upgrades the runtime for you. A container that defines its own Node version is yours to patch. One attendee's team updated about 40 repos in a single day.

Cheap attackers, liability and who's exposed

With hacking agents this inexpensive, almost any custom software can be broken given enough time and money, and the amount of generated code keeps growing the surface. Someone pointed out that this was always true for state actors. The change is how much more power low-skill attackers have. Strong encryption still holds up, and the bigger worry is lateral movement once an attacker is inside a network.

Liability and insurance are a big unanswered question for agents acting on people's accounts. One attendee compared the coming layer of agents attacking and defending to the nanobot battles in Neal Stephenson's The Diamond Age.

Others are more optimistic. White hats have been using agents for a couple of years, security investment will ramp up, and practices will improve to match. Some expect development to slow in places, since more changes will need agent-driven pen testing before they ship.

Supply chains and shifting foundations

The foundations under our stacks are moving quickly, from Bun's port from Zig to Rust to the TypeScript compiler's port to Go. The speed gains are real, but swapping one package can put a weeks-old codebase under a mature site. The XZ Utils backdoor, planted slowly over about two years, is the kind of risk that grows during big migrations.

Supply chain attacks worry several of us most, since many widely used projects are run by one person or a small team and contributing has never been easier. Some wonder whether teams will consolidate around older, thoroughly vetted languages like Java and C#. We don't think old automatically means secure. On Elixir, the smaller ecosystem has fewer white hats looking for problems, though the maintainers of the BEAM runtime keep up well.

What you can do now

• Add a CI check that compares your resolved dependencies against OSV, the open source vulnerability database. One team's check flagged a problem within a couple of weeks.
• Don't rely on Dependabot alone. It doesn't surface Erlang and Elixir vulnerabilities in its UI, so run checks in CI.
• Pin dependencies and GitHub Actions to full commit SHAs, and wait a few weeks before adopting new releases. Even security tools get compromised, as Trivy was this year.
• Watch advisories and check your codebases against them. One team is connecting that to an autonomous agent that opens the patches.
• Try deepsec, Vercel Labs' open-source harness that uses coding agents to find vulnerabilities.
• Use short-lived, tightly scoped credentials such as npm trusted publishing, and expect AI to make social engineering emails more convincing.
• On a home server, move SSH off port 22, disable root login, and use authorized keys.

To close, we asked whether you can run several harnesses together. Yes, through one interface that calls multiple harnesses underneath, such as Conductor. The key is a single entry point for configuring MCP connections and loading skills.

The Econify.AI Roundtable meets Fridays at 11am ET / 4pm UK. Details and registration are at econify.com/ai-community.