When industry leaders gathered at SecDays France to discuss agentic AI, the question wasn't whether this technology would transform business operations—it was whether organizations could harness its power without falling victim to an unprecedented security nightmare.

Beyond Workflows: Understanding the Agentic Revolution

Unlike the deterministic workflows that have dominated enterprise automation, agentic AI represents a paradigm shift toward autonomous decision-making. As Noé Achache, Head of AI Engineering at Theodo, explained, traditional LLM implementations follow predictable paths—call an API, process the response, branch left or right based on predetermined logic. Agentic systems, however, operate with a fundamentally different philosophy: give them access to tools, data, and objectives, then let them determine the optimal path forward.

This distinction becomes crucial when considering protocols like MCP (Model Connection Protocol) and A2A (Agent-to-Agent), which have emerged as the connective tissue enabling this new ecosystem.

MCP, launched by Anthropic in late November 2024, has already spawned thousands of community-built servers in just six months—a testament to both its utility and the voracious appetite for AI automation across industries.

A Look Into the Secrets of MCP: The New Secret Leak Source

MCP rapidly enhances AI capabilities but introduces security challenges through its distributed architecture. Especially, the distributed nature of MCP requires a lot of NHIs and their secrets. Our research shows that MCP is a new source of leaks that already discloses real-world secrets.

GitGuardian Blog - Take Control of Your Secrets SecurityGaetan Ferry

The democratization aspect cannot be overstated. Where previous automation technologies required specialized technical knowledge, agentic AI tools are accessible to anyone who can articulate their needs in natural language. This accessibility is driving adoption rates that dwarf previous technology cycles, creating what panelists described as both an unprecedented opportunity and a looming crisis.

The Traffic Surge: When Bots Become the New Normal

Gilles Walbrou from DataDome brought sobering statistics to the discussion. His company, which has been tracking bot traffic since 2015, has witnessed an explosion in AI-driven web traffic. What started as a manageable stream of automated requests has evolved into a torrent that's reshaping how websites think about security and access control.

The challenge extends beyond simple volume. Traditional bot detection relied on behavioral patterns that remained relatively static over time. AI agents, however, can adapt, learn, and modify their approaches in real-time. This creates an arms race between detection systems and increasingly sophisticated automated actors, some benevolent, others decidedly not.

Notably, industry responses are starting to emerge: Cloudflare just launched a marketplace that allows websites to charge AI bots for scraping their content, an initiative that could shift both the economics and the security dynamics of bot-driven traffic (read more on TechCrunch). Yet as Walbrou emphasized, the fundamental challenge remains: distinguishing between legitimate automation and malicious activity in an environment where the lines are increasingly blurred.

The Identity Crisis: When Machines Outnumber Humans

Perhaps the most profound security implication discussed was the explosion of non-human identities (NHIs) within enterprise environments. Arnault Chazareix drew a compelling parallel to the microservices revolution that began fifteen years ago, noting how each new service required its own authentication mechanisms and access controls. Agentic AI is triggering a similar multiplication, but at an exponentially greater scale.

Every AI agent requires credentials to function. Every MCP server needs authentication tokens. Every automated workflow demands access to the systems it manipulates. The result is a sprawling landscape of machine identities that can quickly outnumber human users by orders of magnitude, creating what security professionals are calling "secret sprawl"—a distributed ecosystem of credentials that's increasingly difficult to inventory, monitor, and secure.

Real-World Attack Chains Involving NHIs: Lessons from SecDays France

Based on the talk “The Explosion of Machine Identities: Are You Aware of This Silent Threat?” at SecDays France 2025.

GitGuardian Blog - Take Control of Your Secrets SecurityAnna Nabiullina

This proliferation isn't merely a scaling problem; it's a fundamental shift in the attack surface. Unlike human identities, which follow predictable patterns of creation, modification, and deactivation, machine identities often exist in a perpetual state of ambiguity. They're created for specific projects, forgotten when those projects evolve, and left dormant with their original permissions intact.

The Democratization Dilemma: When Everyone's a Developer

One of the most fascinating aspects of the current AI revolution is how it's democratizing software creation through what the panel termed "vibe coding."

Automated Guard Rails for Vibe Coding

Vibe coding might sound like a trendy term, but it’s really just developing software without automated checks and quality gates. Traditional engineering disciplines have always relied on safety measures and quality controls, so vibe coding should be no different in my honest opinion.

GitGuardian Blog - Take Control of Your Secrets SecurityAndy Rea

This phenomenon, where non-technical users can generate functional code through natural language prompts, represents both a tremendous productivity opportunity and a significant security risk.

As the discussion revealed, tools like GitHub Copilot can inadvertently generate insecure code, embedding vulnerabilities or exposing secrets in ways that might not be immediately apparent to users who lack deep security training. The panel highlighted emerging threats like "slop squatting," where malicious actors create packages with names commonly hallucinated by AI systems, waiting for unsuspecting users to download and execute compromised code.

GitHub Copilot Security and Privacy Concerns: Understanding the Risks and Best Practices

Worried about GitHub Copilot’s security and privacy concerns? Learn about potential risks and best practices to protect yourself and your organization while leveraging AI.

GitGuardian Blog - Take Control of Your Secrets SecurityDwayne McDaniel

The challenge is compounded by the speed at which AI-generated solutions can move from proof-of-concept to production. Traditional security review processes, designed for human-paced development cycles, struggle to keep up with the velocity of AI-assisted creation. This creates a dangerous gap where insecure code can proliferate faster than security teams can identify and remediate it.

Building Guardrails for the Agentic Future

Despite these challenges, the panel remained optimistic about solutions emerging to address agentic AI's security implications. A key trend they identified is the centralization of AI access through internal proxies and gateways—what some organizations are implementing as "LiteLLM" architectures that route all AI interactions through monitored, controlled channels.

Organizations looking to implement robust AI security frameworks can now reference the newly released OWASP AI Testing Guide, which provides a comprehensive methodology for systematically assessing AI systems across various dimensions including adversarial robustness, privacy, fairness, and governance (learn more about the guide).

These centralized approaches enable several critical security capabilities: mandatory guardrails for all AI interactions, comprehensive audit trails, automated detection of personally identifiable information (PII), and the ability to enforce organizational policies consistently across all AI use cases. The upcoming EU AI Act, which takes effect in August, will further incentivize such approaches by requiring high-risk AI providers to analyze, understand, and test their systems comprehensively.

The panel also discussed the emergence of specialized security agents—AI systems designed specifically to review and improve the security posture of other AI-generated content. This meta-approach to AI security represents a fascinating evolution where the technology itself becomes part of the solution to its own risks.

The Path Forward: Security in the Age of Agents

As the SecDays France roundtable concluded, participants shared their most exciting agentic AI use cases while acknowledging the responsibility that comes with this power. From automating complex business processes to unlocking previously inaccessible unstructured data, the potential applications seem limitless.

Yet the overarching message was clear: agentic AI is not just another technology to be secured—it's a fundamental shift that requires rethinking how we approach identity, access, and risk management. The security community's response must be equally transformative, embracing new methodologies while maintaining the fundamental principles that have guided cybersecurity for decades.

As one panelist aptly summarized: "We are all builders, and we are all becoming security-aware."

In the age of agentic AI, this dual identity isn't just beneficial—it's essential for navigating the extraordinary opportunities and unprecedented risks that lie ahead.

The conversation at SecDays France highlighted that while agentic AI represents uncharted territory, the fundamentals of security remain constant: inventory your assets, monitor continuously, and build governance into every layer of your technology stack. Only by embracing these principles can organizations safely harness the revolutionary potential of autonomous AI systems.