AI systems /
Frontier AI changes the cybersecurity baseline
Much of the software businesses depend on was built for a world where careful analysis was slow and expensive.
Frontier AI changes that. Strong coding agents can read unfamiliar code, map integrations and reason through combinations of permissions and edge cases far faster than a person working alone.
They do not create most weaknesses. They make existing weaknesses easier to find.
Old systems were built for human-speed scrutiny
Legacy software is not automatically insecure. Many teams built carefully within the tools and security expectations of their time.
The risk comes from what accumulates afterwards: forgotten admin pages, old roles, shared service accounts, stale dependencies and integrations that still have production access.
Some of those weaknesses persist because finding a path through them takes time. An attacker needs to understand the code, the business logic and the assumptions that have drifted since the system was built.
Agent-assisted analysis reduces that cost. It can connect a permissive API, an exposed log and an old role into an attack path a human review missed.
The business experiences that as a breach. The system experiences it as a chain of reasoning.
Use the same capability defensively
This does not replace existing security practice. Identity controls, patching, monitoring, secure development and incident response still matter.
It raises the standard for reviewing systems that hold customer accounts, payments, personal information, internal administration or operational infrastructure.
Start with a short list of the systems where failure would matter. Then use strong tools and security specialists to answer practical questions:
- Which roles, endpoints and integrations still have access they no longer need?
- Which parts of the codebase would the current team struggle to explain under pressure?
- Can the authentication and authorisation paths survive targeted testing?
- Which logs, backups and exports expose sensitive data?
- How quickly can a confirmed finding become a patch, test or permission change?
An annual audit remains useful. It is not a measure of whether a system can withstand agent-speed analysis between audits.
Keep people responsible for consequential changes
AI can help map a system, propose tests and identify suspicious paths. It should not receive open authority to change production systems because it found something interesting.
Keep the investigation, evidence and proposed fix visible. Require a human to approve changes that affect data, money, access or availability. Turn confirmed findings into tests so the weakness does not return in a later release.
The value is not a model that replaces security judgement. It is a security team that can inspect more of the real system before an attacker does.
The rule
Do not assume old software is safe because it survived the old internet.
Test important systems against agent-speed reasoning, then fix what the review finds.