Harness, Socure and Operant AI: the latest vendor and product news

Michiel Willems
September 2, 2026

Your weekly roundup of the most important QA, software testing and software delivery launches for banks and financial services firms, alongside investments, partnerships and other vendor news.

This week, Harness targets the growing volume of AI-generated code, Socure raises $156 million and acquires Fravity, while Operant AI and TRUSTNOW launch platforms intended to control autonomous agents.

Elsewhere, Anthropic resumes external cybersecurity testing of its models, Coforge introduces an agentic development framework and QuantumX moves its quantum simulation platform beyond beta testing.

Harness targets AI-generated code

Harness has launched a code repository and AI Code Review capabilities designed for development environments producing increasing volumes of AI-generated software.

The repository includes permissions intended specifically for AI agents, alongside automated reviews and migration tools for teams moving from existing Git providers. Harness said its AI Code Review can examine proposed changes and provide feedback within development workflows.

The launch reflects a growing problem for software delivery teams: AI coding assistants can increase the volume of changes awaiting review faster than conventional quality controls can process them.

For banks, that makes traceability particularly important. Development teams need to establish which code was produced or altered by an agent, which tests were run, who approved the change and whether security or regulatory controls were satisfied before deployment.

Automated review may help teams manage the additional volume, but financial institutions will still need to validate the reviewer itself and retain human approval for higher-risk changes.

Socure raises $156m and buys Fravity

Identity verification specialist Socure has raised $156 million through an extension of its Series E financing, valuing the business at $5.2 billion.

The investment was led by Summit Partners, with participation from Goldman Sachs Alternatives, Wells Fargo, DocuSign and other investors.

Socure has also acquired Fravity, an agentic operations platform focused on automating fraud, risk and compliance work. Financial terms of the acquisition were not disclosed.

Fravity’s technology will be incorporated into Socure’s RiskOS platform and offered as RiskOS Agents.

The combined system is intended to automate elements of fraud and compliance operations rather than limiting AI to generating recommendations for human users.

Socure serves more than 3,000 customers across over 190 countries, including financial services firms.

The addition of autonomous agents creates an important assurance question for banks using the platform. Testing will need to establish what decisions or actions an agent is permitted to take, when cases are escalated to a person and whether the system produces consistent outcomes across different customer and fraud scenarios.

Operant AI launches agent firewall

Operant AI has launched a Semantic Firewall designed to assess the intent of autonomous agents and enforce enterprise policies while they operate.

The platform examines prompts, model responses, commands, tool calls and data movement. It can then allow, block or redact an action before it is completed.

Operant has added separate controls covering agent access to tools, code and sensitive data. A scope guard is intended to determine whether an agent’s instructions and proposed actions remain consistent with its authorised purpose.

The company has also introduced live protection for AI activity conducted through authenticated browser sessions, wider coverage for Claude-based tools and a token meter for monitoring usage and enforcing budget limits.

The launch is particularly relevant to banks exploring agents that can access customer information, update records or connect with internal applications.

Testing cannot focus only on whether the agent completes its assigned task. Firms must also validate that the firewall detects indirect prompt injection, unauthorised tool use, attempts to move sensitive data and sequences of individually legitimate actions that produce an unacceptable result when combined.

TRUSTNOW adds real-time agent governance

Deep-tech startup TRUSTNOW has launched EKAM, an AI governance platform intended to give enterprises visibility and control over autonomous agents.

The platform can identify authorised and potentially unmanaged agents, assign cryptographic identities and enforce least-privilege access. Policies can be used to allow, block or escalate sensitive actions in real time.

EKAM also includes a kill switch and produces an auditable record of agent activity, authorisations and actions. It supports self-hosted, hybrid and air-gapped deployments and is designed to operate independently of the underlying AI model.

“When AI moves from answering to acting, governance has to be real-time and can no longer be an afterthought,” said TRUSTNOW founder and CEO Raguram Gopalan.

For regulated financial institutions, the audit trail may be as important as the immediate control. Banks will need evidence showing what authority an agent held at a particular time, which data or systems it accessed and why an action was permitted or stopped.

Anthropic resumes external security testing

Anthropic has resumed external cybersecurity testing of its AI models after introducing additional safeguards.

The company temporarily halted the evaluations following incidents in which Claude models accessed company systems during cybersecurity tests.

The resumption demonstrates the unusual challenge involved in evaluating increasingly capable AI. A security test must give a model enough access and freedom to reveal its capabilities without allowing the exercise itself to create an unacceptable risk.

The issue is relevant to financial institutions conducting adversarial evaluations of models and agents. Test environments need strong isolation, realistic but controlled data, clear stopping conditions and monitoring capable of detecting when a model moves beyond the intended scope of an exercise.

Results must also be reproducible enough to determine whether new safeguards have corrected the behaviour or merely changed the circumstances in which it appears.

Coforge builds agentic development framework

Coforge has introduced an AI adoption framework designed to help enterprises expand agentic software development beyond initial pilot projects.

The framework combines Kotter’s Change Model, the Prosci ADKAR methodology and Operational Change Management. It covers workforce readiness, development operating models, governance and the measurement of AI-tool adoption.

Coforge said the approach addresses changes to code review, testing, requirements management, quality controls and team responsibilities as AI agents take a larger role in software development.

“The most successful organizations treat agentic development as an enterprise transformation initiative rather than a technology project,” said Anup Nair, chief AI commercial officer at Coforge.

For banks, adopting AI coding agents changes more than the speed at which software is produced. Testing responsibilities, approval gates and segregation of duties may all need to be reconsidered when an agent can interpret requirements, write code and propose changes across several stages of delivery.

QuantumX takes simulation platform beyond beta

QuantumX has launched a hosted version of its EntangleX quantum simulation platform following external partner beta testing that began in July.

The browser-based platform is intended to make quantum simulation tools accessible without requiring users to maintain specialist local infrastructure.

Although practical quantum computing remains at an early stage, financial institutions are already exploring potential applications across optimisation, modelling and risk analysis.

For testing teams, simulation platforms provide a way to evaluate algorithms and integration approaches before suitable quantum hardware is widely available. Results will still require careful validation, particularly when classical approximations, simulated noise and experimental quantum workflows are compared.