The AI agent security market is consolidating. The evidence gap isn't.

Two acquisitions and two large funding rounds show how quickly the AI security category is developing.

Check Point agreed to acquire Lakera, whose platform focuses on security for generative and agentic AI applications. SentinelOne agreed to acquire Prompt Security, which works on security for generative AI usage and agentic systems. WitnessAI announced $58 million in strategic funding alongside expanded agentic AI security capabilities. HappyRobot, whose expansion plans include banking and other enterprise workflows, raised $150 million at a reported $1.2 billion valuation.

The details differ, but the direction is clear: significant capital is moving toward systems that help enterprises detect and control risks in AI applications and agents.

Prompt injection, data leakage, jailbreaks, and misuse are real problems. Products designed to detect and block those threats are becoming more capable, more visible, and better funded.

A different question is receiving less attention: how do you prove, after the fact, that a specific agent action was authorized?

Two different questions

"Is this agent under attack?" and "Was this agent allowed to do this?" sound similar. They are different questions.

The first is primarily a security question. It involves checking behavior against attack patterns, suspicious inputs, data-exfiltration attempts, or other indicators of misuse.

The second is a governance and compliance question. It involves checking an action against the organization's policies, recording the decision, and preserving enough context to explain that decision later.

The examples above are primarily positioned around detection, protection, monitoring, and control of AI usage and agent behavior. Those capabilities are valuable. They do not automatically produce the same evidence as a record showing which policy was evaluated, what verdict was reached, whether the action was blocked or allowed, and whether human approval was required.

A security event log and an authorization record can sit next to each other, but they serve different purposes. One helps a security team investigate an attack or misuse. The other helps an auditor, risk owner, or regulator reconstruct a control decision.

Why the evidence layer matters

As agents move from drafting and summarizing toward taking actions in enterprise systems, the consequences of those actions become more important. An agent may access records, call tools, update systems, communicate externally, or participate in a regulated workflow. Each action can carry permissions, data-handling requirements, and approval rules that exist independently of the model producing the action.

The regulatory timeline is also becoming more concrete, although individual obligations depend on the system, role, and jurisdiction.

Article 50 of the EU AI Act applies from August 2, 2026 and establishes transparency obligations for certain AI systems, including requirements concerning disclosure of AI interaction and the marking or labelling of certain AI-generated content. Article 50 should not be described as a general requirement for a per-action authorization log. That is a different claim.

The evidence question is closer to the broader record-keeping and governance problem addressed in our existing work on Article 12. Organizations need to understand what their systems did, which controls were applied, and whether they can produce reliable evidence when an internal reviewer, auditor, customer, or regulator asks for it.

Regional regulation is moving in the same direction. DIFC Regulation 10 covers autonomous and semi-autonomous systems in the DIFC, and Qatar Central Bank's AI Guideline sets governance expectations for licensed institutions. The exact obligations depend on the system, role, and jurisdiction, but the operational question is similar: can an institution explain how an AI system was governed and what happened when it acted?

A security tool that blocks a prompt injection attempt produces a security log. It does not automatically produce a record showing which policy a specific agent action was evaluated against, what the verdict was, and whether a human needed to approve it. One artifact is written for a SOC analyst. The other is written for an auditor.

What we are building toward

That second artifact is what we build at Gateplex. We intercept an agent's action before it executes, evaluate it against configurable rules, and write a tamper-evident, hash-chained record of that decision as it happens.

Gateplex currently supports ten rule types: Keyword Block, Keyword Flag, Rate Limiter, API Scope Control, Regex Pattern Match, Output Length Limit, Prompt Injection Detector, Repetition Detector, Spend Limit, and PII Shield. Where a workflow requires it, a flagged action can enter an approval queue for human review.

In our implementation, evaluation is deterministic, with no model in the decision path, so the same input produces the same verdict every time. The output is not a security log describing an attack that was stopped. It is a compliance record describing a decision that was made, and why.

We do not expect to have this layer to ourselves for long. Governance platforms built around documentation and periodic assessment are moving toward real-time capability. Infrastructure gateways already handle identity and cost control at the point where agents act, and evidence generation is a natural next step for them too. The gap exists today. It will not stay this open.

For now, the pattern holds. The market is getting very good, very quickly, at stopping AI agents from doing the wrong thing in the moment. Proving what an agent was allowed to do in the first place is getting far less attention, and it is the part regulated industries will need answered first.

Sources

• European Commission, "Transparency obligations under Article 50 of the AI Act": digital-strategy.ec.europa.eu
• SentinelOne, "SentinelOne to Acquire Prompt Security to Advance GenAI Security and Agent Security Strategy," August 5, 2025
• Check Point's acquisition of Lakera, reported by CTech and SecurityWeek, September 2025
• WitnessAI's $58 million strategic funding announcement, reported by PR Newswire, January 2026
• HappyRobot's $150 million financing, reported by SiliconANGLE, August 4, 2026: siliconangle.com
• DIFC Regulation 10 on autonomous and semi-autonomous systems
• Qatar Central Bank AI Guideline for licensed institutions
• Gateplex, "EU AI Act Article 12 evidence"
• Gateplex, "FLAG verdicts and human approval"