Replit's Amjad Masad: AI Safety Is a Cyber Security Problem

Key points

Key takeaways from The Information interview with Replit co-founder and CEO Amjad Masad (September 2026):

Steadier scaling curve. New frontier models now arrive every two to three weeks, with benchmarks such as ARC-AGI saturating in the low 80s. Per-token prices and tokens-per-task are both falling, so the overall cost of using models keeps dropping even as total token usage rises.

Code generation is commoditising. Masad argues generating code is becoming a commodity — every major assistant, from Muse to Instinct to Claude, now has coding capabilities — so Replit differentiates on the harness around the model: sandboxing, database, memory and context, identity and authorisation, network security, deployment and managing agents at scale, including secure MCP connections.

The AI safety debate misses the cyber security story. Three of the four recent breaches at major labs (OpenAI, Hugging Face, Meta, Anthropic and Google's Gemini) involved the same sandbox contractor, Invisible, via a misconfiguration — a story Masad says is rarely covered. He points out that agents have hacked RubyGems, Hugging Face and PyPI, which he describes as hacking core internet infrastructure.

Enforce existing laws, don't pace the frontier. Masad argues AI safety policy should focus on enforcing existing software laws — if an agent commits a crime, someone should be liable — rather than introducing new frontier-pacing frameworks. He warns of regulatory capture and says it is premature to ask governments to decide how fast labs may build.

Labs should slow themselves down. Reading the incident logs, Masad says, showed no monitoring and no "safety red button" with which to shut systems down. Labs should pace themselves and harden their systems before shipping further — otherwise they risk criminal liability.

Rival red-teaming is welcome. He endorses the idea of frontier labs testing each other's models with adversarial threat techniques — such as the reported deal between OpenAI and a rival lab to probe each other's systems — as the kind of accountability configuration the industry needs.

Europe is "eating the bait". Masad says European and UK politicians are taking the big labs' framing of the safety debate much more readily than the US administration, which he says has tech people who understand incentives and the risk of regulatory capture. He cites Europe and the UK recently banning superintelligence as an example.

London's realist-optimist approach. After opening an office in London, Masad met Mayor Sadiq Khan and the Prime Minister's office. Contrasting New York City's ban on AI in schools with London's openness, Replit is collaborating with the city on training and upskilling programmes — Khan describes himself as a "realist optimist", a stance Masad prefers to pure accelerationism.

Read more: The Information · Watch on YouTube