Security Funding Surges as Traditional Cyber Defences Prove Inadequate for AI Era

Legacy Architectures Meet Their Limit

Security companies are pulling in investment rounds that would have seemed extravagant eighteen months ago. Instinct and Simile - two startups building defences explicitly designed for generative AI workloads - have each closed financing in excess of $100 million, at valuations their founders acknowledge would not have been plausible before large language models entered production environments at scale.

The capital influx reflects a broader recognition among limited partners and fund managers: the perimeter-based security model that has underpinned enterprise IT for two decades is structurally unsuited to a world in which software writes itself, agents operate with delegation, and attack surfaces expand faster than security teams can map them.

Public equity markets are registering the same shift. Cybersecurity indices have climbed steadily over the past four quarters, outpacing broader technology benchmarks even as concerns about AI safety and the prospect of rogue autonomous agents dominate policy discussions from Brussels to Washington to Seoul.

What Changed Beneath the Surface

At Opentechwire, we have tracked security spend across Asia-Pacific enterprises for three years. Until mid-2025, most chief information security officers treated generative AI as a data-leakage problem - employees pasting proprietary code into ChatGPT, sensitive customer records ending up in training corpora. The countermeasures were procedural: acceptable-use policies, endpoint monitoring, API gateways that log prompts.

That framing broke in late 2025, when two events converged. First, multiple enterprises began deploying agents with write access to production databases, CRM systems and financial ledgers, delegating decisions that had always required human sign-off. Second, researchers demonstrated that adversarial inputs could reliably manipulate agent behaviour in ways that static code analysis and signature-based intrusion detection simply do not catch.

The result is an environment in which the threat model has fundamentally changed. Traditional security assumes a relatively static codebase, a defined network perimeter and humans in the loop for high-stakes actions. AI-native environments feature none of those properties. Code is generated on-demand, perimeters dissolve when agents call third-party APIs autonomously, and the loop tightens to the point where human oversight becomes a performance bottleneck rather than a safety check.

The Instinct and Simile Phenomenon

Instinct raised $120 million in a Series B round that closed in August 2026, according to the company. The startup's core product monitors agent behaviour in real time, flagging actions that deviate from learned patterns of legitimate delegation. Instead of scanning for malware signatures, Instinct's platform builds probabilistic models of what a given agent should be doing in a given context - which APIs it should call, which database tables it should query, which external services it should invoke - and raises alerts when observed behaviour falls outside those bounds.

Simile, which announced a $150 million Series C in early September 2026, takes a different approach. The company embeds lightweight verification layers directly into the inference pipeline, checking each model output against a policy graph before the agent acts. If an output violates a constraint - attempting to transfer funds above a threshold, modifying a configuration file flagged as critical, or exfiltrating data to an unapproved endpoint - the verification layer blocks execution and logs the attempt.

Both companies are growing at triple-digit year-on-year rates, according to figures they have shared with investors. Neither would have existed five years ago, because the problem they solve did not exist at enterprise scale.

Why Traditional Vendors Struggle to Adapt

Incumbent security vendors - the firms that built fortunes on firewalls, endpoint agents and security information and event management platforms - face a structural disadvantage. Their products were architected for a world in which software changes slowly, in discrete releases, and runs on infrastructure the enterprise controls.

Generative AI inverts those assumptions. Models update continuously through fine-tuning. Agents operate across organisational boundaries, calling APIs hosted by third parties whose security posture the enterprise cannot audit. The artefacts that traditional tools inspect - compiled binaries, network packets, file hashes - are less useful when the relevant "code" is a prompt, a set of embeddings, or a policy expressed in natural language.

Retrofitting legacy platforms to handle these dynamics is proving difficult. Several established vendors have announced AI security modules over the past year, but adoption has been tepid. Security leaders we have spoken with in Singapore, Tokyo and Sydney report that bolted-on features often generate too many false positives, lack integration with the agent frameworks developers actually use, or introduce latency that makes real-time intervention impractical.

The funding flowing to Instinct, Simile and a dozen smaller startups suggests investors believe the incumbents will not close that gap quickly enough. The window for native solutions is open, and capital is moving accordingly.

The Regulatory Tailwind

Policy is amplifying the investment case. The European Union's AI Act, which entered force in stages throughout 2025 and 2026, imposes liability on deployers of high-risk AI systems if those systems cause harm through inadequate safeguards. Singapore's Monetary Authority has issued similar guidance for financial institutions using autonomous agents in trading, lending and fraud detection. South Korea's Personal Information Protection Commission released draft rules in July 2026 that would require real-time auditing of any agent with access to citizen data.

Compliance with these frameworks is not optional, and it is not something general-purpose security tools were designed to deliver. Enterprises need platforms that can produce audit trails showing which agent took which action, under what policy, with what human oversight - and they need those trails in a format regulators will accept.

Instinct and Simile have both hired former regulators to shape their product road maps. That is not coincidental. The startups are building to a compliance standard that is still being written, betting that early alignment with regulatory expectations will create switching costs once enterprises commit.

Open Questions on Efficacy

The capital flooding into AI security does not guarantee that the solutions being funded will work at scale. Agent behaviour is probabilistic, which makes it inherently harder to verify than deterministic code. A model that behaves safely in testing may produce an unexpected output in production when confronted with an edge case its designers never anticipated. Verification layers add latency, and latency can render an agent useless for time-sensitive tasks.

There is also a cat-and-mouse dynamic that no amount of venture funding can eliminate. Adversaries will probe these new defences, find weaknesses and adapt. The startups raising nine-figure rounds today are betting they can iterate faster than attackers - a bet that has not always paid off in earlier waves of security innovation.

What the Funding Wave Signals

The scale of investment in AI-native security reflects a broader reckoning. Enterprises, investors and regulators have concluded that the old model - perimeter defences, signature-based detection, human-in-the-loop controls - cannot contain the risks that autonomous agents introduce.

Whether the new generation of startups can deliver on that promise remains to be seen. But the capital is committed, the regulatory pressure is building, and the problem is not going away. For security vendors, the question is no longer whether AI requires a new approach. It is whether they can build that approach fast enough to matter.