Overview of AI broke the cybersecurity world. Now what?
In this Decoder episode, Nilay Patel talks with cybersecurity veteran Alex Stamos, chief security officer at Cognition and a Stanford lecturer, about how AI is reshaping the internet’s threat landscape. Stamos argues that the traditional cybersecurity model is effectively broken: open-weight AI models now let attackers automate much of the “kill chain,” find vulnerabilities in real time, and launch highly scalable attacks at a fraction of the previous cost. The result is a new era of constant breaches, faster exploitation, and a widening gap between cheap offense and expensive defense.
The Big Thesis: AI Has Changed Cyber Economics
Stamos’ core argument is that AI has upended the assumptions that sustained the cybersecurity industry for the last two decades:
- Attackers no longer need rare elite talent to carry out sophisticated operations.
- Open-source / open-weight models lower the barrier to entry for “script kiddies” and professional threat actors alike.
- The old split between research teams and operations teams is collapsing.
- Defensive organizations are still expensive and slow, while attackers can scale quickly and cheaply.
He describes this as the start of a “dark age” for cybersecurity—not the end of the world, but a period where defenders will struggle to keep up until systems, tooling, and incentives catch up.
What’s Happening Right Now
Stamos points to a string of recent incidents as evidence that this shift is already underway:
- Large breaches across AI-related companies and cloud infrastructure providers
- The Cursor incident, where stolen API keys helped chain one breach into many more
- The Binance hack, which he says was largely a North Korean operation focused on stealing Bitcoin
- The use of tools like “Artex,” which he describes as enabling rapid, automated intrusion against many organizations
His view is that these are not isolated events—they are signs of a more permanent, continuous breach environment.
Why AI Makes Attacks So Much Harder to Defend Against
Stamos explains that AI changes cyber operations in several ways:
Faster vulnerability discovery
AI can now help attackers:
- analyze targets
- find bugs
- generate exploits
- deploy them immediately
That compresses what used to be a longer, human-driven process into something close to real-time.
Lower barrier to entry
Attackers can now use:
- open-weight models
- consumer GPUs / Macs
- local quantized models
- inexpensive automation
This means smaller groups can now do the work once limited to intelligence agencies or top-tier criminal organizations.
Economics favor offense
He emphasizes a key asymmetry:
- Attacker marginal cost: trending toward zero
- Defender marginal cost: still high
That imbalance, he says, is the central problem.
The Cybersecurity Industry’s Old Model Is Breaking
Stamos says the modern security industry grew out of a specific era:
- Microsoft’s “trustworthy computing” shift
- responsible disclosure culture
- bug bounties
- consulting and incident response firms
- a premium on rare elite skill
That world assumed:
- elite hackers were scarce
- operational attack teams were limited
- defense could keep up through patching and process
AI breaks those assumptions. It makes high-end capability widely available and undermines the old skill hierarchy.
What Defenders Need to Do
Stamos argues that simply finding and fixing bugs is not enough. Defense has to become cheaper and more automated across the board.
1. Automate software remediation
He says the most effective near-term use of AI is:
- scanning codebases
- finding vulnerabilities
- writing tests
- generating patches
- merging fixes automatically
In some cases, he believes it is cheaper to rebuild old systems than to keep patching them forever.
2. Use AI in security operations
The next major area is SOC work:
- log analysis
- alert triage
- incident investigation
- deciding when to respond and how hard
He argues that humans alone cannot handle the scale of modern telemetry, and that AI systems—especially cheaper, smaller ones—must be used to augment or replace first-tier defense work.
3. Build bounded defensive agents
Stamos is open to defensive AI agents, but only with strict guardrails:
- deterministic limits
- account caps
- escalation paths to humans
- no “let the swarm loose” approach
He is skeptical of loose, autonomous security “swarm” rhetoric and says most real-world attacks are still much more constrained than that.
What Companies Should Do Now
One of the strongest recommendations in the episode is practical and urgent:
- Every company should red team itself immediately.
- If not with actual adversarial tools, then with AI-based red teaming products.
- Companies that don’t test themselves will be breached by attackers doing exactly that.
His message: organizations need to proactively simulate the attack before someone else does.
The Linux Kernel, OpenSSL, and the Need to Rebuild
Stamos makes a provocative argument that patching the foundational stack is no longer enough:
- Linux kernel bugs keep appearing
- OpenSSL and glibc remain major risk surfaces
- Patch Tuesday keeps getting larger
- exploitation timelines are too fast for the old patch-and-pray model
He suggests that companies like Google or Amazon should consider forking and rewriting critical infrastructure in memory-safe, type-safe languages, at least for their own cloud/server environments.
His broader point: the industry should stop assuming it can patch its way through decades of legacy code.
AI, Social Media, and Synthetic Content
The conversation also shifts to the broader internet and AI-generated media.
AI slop on social platforms
Nilay raises concerns about Meta and Instagram being flooded with synthetic content, including deepfakes and low-quality AI-generated “slop.” Stamos agrees that this will become a major issue.
Verifying real content may be better than detecting fake content
He argues that:
- detecting fake content is probabilistic and gets harder over time
- verifying real content can be cryptographically strong
His preference:
- prove authenticity at the capture/source level
- build a chain of custody from camera to platform
- restrict editing to trusted environments
- make real-origin verification the default
He believes a new app or platform could disrupt Instagram by making authenticity central.
The Human Risk: AI as a Tool for Social Engineering
Stamos is especially concerned about “human uplift risk”:
- phishing
- voice scams
- fake kidnap videos
- fake friends / AI companions
- manipulation through realistic synthetic media
He thinks one of the most dangerous trends is AI systems that present themselves as emotionally human or as “friends.” He calls that unethical and says the industry should reject it socially, if not legally.
Where He Thinks Progress Will Show Up
Stamos is not a pure doomer. He says there are signs to look for if things improve:
- fewer major breaches
- fewer wormable CVEs in core infrastructure
- fewer Linux kernel and OpenSSL-style emergency patches
- a decline in Patch Tuesday volume
- better cost-efficiency for defensive tooling
- more code rewritten in safer languages
- more AI-assisted remediation of legacy systems
He would like to see a future where the worst vulnerabilities are less common and the cost balance shifts back toward defense.
Key Takeaways
- AI has dramatically lowered the cost of cyber offense.
- The old cybersecurity industry model is no longer sufficient.
- Defense must become far more automated and far cheaper.
- Companies need active self-red-teaming now, not later.
- Critical infrastructure likely needs to be rewritten, not just patched.
- Synthetic media and AI-generated social content are becoming a parallel security problem.
- Verifying authenticity may be more durable than trying to detect every fake.
Notable Quote
“We are entering a dark age.”
That line captures Stamos’ overall tone: not apocalyptic, but deeply concerned that the internet’s security foundations are entering a prolonged, difficult transition.