From Niche Analytics to Natural Language APIs

AI has been conflated with LLMs recently. This is largely because LLMs brought AI into mass adoption. Previous generation of AIs touched limited surfaces, such as big company that do sentiment analysis, product recommendation, data churns and so on. Not until the LLM era where ChatGPT got released that AI touch most human beyond big corporation use cases. LLM serve as the natural language API endpoint for humans.

The LLM evolution itself is moving fast from a static chatbots into agentic systems. There is RAG to enrich the frozen in time model state. There is MCP that enable agent to connect with different tools and capability. A simple CLI can even replace MCP if agent has the environment to run it. With the capability of reading the internet, documentation, and papers and the capability to execute tools we are facing unprecedented risk to information security.

When AI Agents Go Offensive

In the past, human hackers found vulnerabilities. But not all vulnerabilities are immediately or easily exploitable. AI changed this equation. Autonomous exploit development by frontier AI agents is no longer a hypothetical capability. This dynamic unfolded recently during an accidental breach caused by OpenAI cybersecurity test, which went haywire and broke in to Hugging Face. This is an example of how automated agents found and exploited without human intervention.

The Defender’s Paradox: When Guardrails Backfire

AI sits in fascinating intersection in the power imbalance setup between attackers and defenders. In the case OpenAI test above, Hugging Face tried to do log analysis of the attack using commercial frontier model, but were blocked by safety guardrails. Thanks to the open weight model, they have GLM-5.2 setup that helped them figure out what is going on. The guardrails in the commercial API purpose was to prevent someone malicious for planning an attack but actually affecting a defender that try to stop an attack. This is also one strong example of why AI must not be monopolised and the importance of open weight model as alternative.

The Case for Open-Weights in a Digital-First World

Cybersecurity in the AI era is becoming as perilous as it is fascinating. Specifically and how our life depends so heavily on digitalisation. Most of financial transaction is digital. Critical infrastructure such as power plant, medical system, transportation are running by computer system. As the difficulty on autonomously building exploit decreases, we need to increase the capability on building defense too. The power imbalance on how few big a handful of Big Tech giants mainly in the US with deep pockets hold a monopoly on frontier models. The alternative open weights model available mostly by China AI company provides some relief, although it is not clear also how long it will stay.