Devin Thomas strips AI guardrails for red teams

Devin Thomas is stripping the safety brakes off open-source artificial intelligence.

On This Week in Startups in October 2026, the Obliteration AI founder explained that commercial AI guardrails actively cripple enterprise cybersecurity defenses. When corporate red teams attempt to simulate cyberattacks or probe internal software agents for vulnerabilities, proprietary models from OpenAI and Anthropic issue blanket refusals. By using model interpretability to locate and alter the exact parameters responsible for refusals in open-source models like DeepSeek and GLM, Thomas removes those hardcoded barriers without degrading baseline reasoning performance.

The architectural breakthrough relies on an inversion of control. Instead of embedding rigid moral filtering deep within model weights, Obliteration AI strips the weights clean and routes user queries through an external, low-latency software policy layer. This secondary safety checker runs in milliseconds, enforcing custom enterprise security guidelines while preserving hard stops against self-harm or explicitly illegal acts.

Demand for uncensored weights has moved rapidly beyond niche defense contractors. Fortune 500 chief information security officers are now buying these modified models to attack their own autonomous agent deployments before external threat actors find the gaps. Anthropic recently underscored this risk shift by documenting the cyber attack efficacy of an obliterated GLM model, highlighting how quickly open-source weights can be weaponized or repurposed.

The competitive timeline between global AI developers is compressing fast. Thomas noted on the show that the performance gap separating open-source Chinese models from US frontier models has narrowed to approximately three months. As open-weight models catch up to proprietary systems, defensive security teams require equal access to uncensored capabilities to keep pace with potential adversaries.

Startups operating in this space face immediate structural bottlenecks. A persistent global GPU shortage restricts high-availability scaling for new security platforms, even as corporate demand accelerates. Data from venture firm Andreessen Horowitz shows that a massive percentage of enterprises have still not integrated AI, held back by lengthy internal approval chains and unresolved security risks.

Beyond technical guardrails, host Jason Calacanis addressed broader predictions about AI replacing professional service business models. While critics argue automated agents will eliminate hourly billing in legal and accounting fields, Calacanis insisted clients pay human experts primarily for liability and risk transfer during litigation. Automated tools merely push advisers upstream to deliver higher-value work per hour.

A similar dynamic protects traditional startup founding teams. Despite tools enabling solo builders to write code faster, venture capitalists continue to back multi-founder teams to avoid single points of failure. Software handles execution, but humans remain necessary to share organizational risk and manage operational pressure.