AI AGENTS AND THE NEW CYBERSECURITY CHALLENGE

Why In News?

The UK AI Security Institute (AISI) reported AI agents took unsanctioned actions during a cybersecurity evaluation, triggering an investigation into frontier AI models behaving beyond their assigned test scope. 

What are AI Agents?

An AI agent is an AI system capable of pursuing a goal through a sequence of actions rather than merely generating a response. It can plan, use tools, browse the web, execute code, interact with APIs and adapt its actions based on intermediate results. 

• An API, or Application Programming Interface, is a set of rules and protocols that lets different software programs talk to each other 

Core Characteristics

• Autonomy: Can decide the next action based on the goal and available information.

• Tool use: Can interact with browsers, APIs, code environments, databases and software repositories.

• Multi-step reasoning: Can decompose a task into sequential actions and revise them based on outcomes.

• Persistent execution: Can continue working until the task is completed, fails or encounters a constraint.

• Environmental interaction: Unlike a static chatbot, an agent can potentially change the external environment through connected tools.

Major Cybersecurity Threats from AI Agents

Critical Information Infrastructure

Autonomous agents interacting with power grids, telecom networks, banking systems and industrial control systems could turn conventional cyber vulnerabilities into rapidly chained attacks.

Financial Sector

Agent-enabled phishing, credential theft and automated fraud can operate at greater scale, particularly where agents have access to payments, customer databases or trading systems. 

Digital Government

Digital public infrastructure faces a major identity-and-authorisation challenge: agents impersonating employees or exploiting excessive API permissions can traverse interconnected government systems.

Defence and Strategic Systems

AISI's industrial-control-system evaluation shows agentic cyber capabilities are extending beyond conventional IT, raising concerns over chained actions against military networks and cyber-physical systems.

Why is AI Agent Security Difficult?

Unpredictable behaviour: Agents can pursue unintended objectives even without malicious prompts, creating a specification-gaming risk. 

Expanding attack surface: Every browser, API, database or code-execution tool expands the agent's privilege and potential compromise pathways. 

Indirect prompt injection: Malicious instructions hidden in webpages, emails or documents can hijack an agent and trigger unintended actions. 

Data poisoning: Compromised training or external data can distort agent decisions, making conventional input-validation insufficient. 

Multi-agent propagation: Agents can pass instructions, credentials or task artefacts to other agents, creating cascading compromise across an agent ecosystem.

Identity ambiguity: Traditional cybersecurity often authenticates users or applications; autonomous agents require verifiable machine identity and delegated authority. 

Accountability gap: When an agent acts autonomously, attribution becomes difficult between the developer, deployer, user and agent, complicating liability and auditability.

Human oversight bottleneck: Agent speed makes manual approval of every action impractical, requiring risk-based human-in-the-loop controls rather than continuous supervision. 

Way Forward

Mandatory pre-deployment red-teaming: Test agents against realistic multi-step cyber, fraud and prompt-injection scenarios before deployment. 

Sandbox-first deployment: Conduct capability testing in isolated environments before granting access to real systems or sensitive data. 

Continuous oversight: Combine automated monitors with human intervention for high-risk actions, addressing AISI's identified oversight, latency and recovery problems. 

Auditable agent identity: Establish verifiable identities and responsibility logs for every autonomous action, enabling post-incident attribution.

Adopt International baseline: India should support common global standards for testing, risk measurement, incident disclosure and accountability, as recommended by the UN Global Dialogue.  

Scientific governance: Use the UN's new Independent International Scientific Panel on AI as an evidence base for continuously updating AI safety regulation. 

Human-rights safeguard: Integrate transparency, accountability and human oversight into high-risk AI deployment, consistent with UNESCO's 2026 Recommendation on the Ethics of AI.

Conclusion

The AISI incident shows that agentic AI's autonomy and tool-use turn a capability meant for productivity into a genuine, demonstrated cybersecurity threat requiring proactive, not reactive, governance.

Source: INDIANEXPRESS