The Double-Edged Sword of Agentic AI: Cybersecurity Experts Demand Urgent Governance and Strategy to Counter Autonomous Threats

As agentic artificial intelligence models transition from theoretical frameworks to operational tools, the global cybersecurity landscape is entering a period of unprecedented volatility. These advanced systems, characterized by their ability to operate autonomously, plan multi-step tasks, and utilize digital tools without constant human intervention, represent a dual-use technology of significant concern. While cybersecurity professionals laud the capacity of agentic AI to sift through petabytes of data at speeds impossible for human analysts, a growing chorus of experts warns that these same capabilities are being weaponized by sophisticated threat actors. In a high-stakes gathering at the Berkman Klein Center for Internet & Society at Harvard University, a panel of leading voices in technology, law, and policy reached a stark consensus: the window for establishing a regulatory framework for agentic AI is closing rapidly, and the consequences of inaction could jeopardize personal privacy, global economic stability, and national security.

The Evolution of the Threat: From Chatbots to Autonomous Agents

The transition from generative AI, which primarily produces text and images, to agentic AI marks a fundamental shift in the digital threat model. Agentic systems do not merely respond to prompts; they pursue objectives. This autonomy allows them to navigate complex networks, identify vulnerabilities, and execute exploits with minimal oversight. According to recent data from IBM, the impact of this evolution is already visible. A comprehensive 2026 study by the firm revealed that cyberattacks targeting public-facing software and systems applications—many of which integrated AI-driven automation—surged by 44 percent year-over-year.

The danger was underscored by a high-profile security breach in November involving Anthropic, the developer behind the Claude AI ecosystem and the Claude Code assistant. In this instance, attackers utilized their own specialized AI models to perform automated reconnaissance on Anthropic’s source code. By scanning for subtle architectural weaknesses that human auditors might overlook, the attackers were able to extract and publish sensitive internal workings of the system. This incident serves as a primary example of "AI-on-AI" warfare, where the speed of the attack outpaces traditional defensive measures.

James Mickens, the Gordon McKay Professor of Computer Science at Harvard University, highlighted the inherent disadvantage faced by defenders in this new era. "The unfortunate thing is that the bad people only have to win once in some sense, whereas the defenders have to win all the time," Mickens noted during the panel. He emphasized that agentic AI compresses the "OODA loop" (Observe, Orient, Decide, Act) to such an extent that human-in-the-loop defense strategies are becoming increasingly obsolete.

The Erosion of Social Engineering Detectors

One of the most immediate and pervasive impacts of agentic AI is the professionalization of phishing and social engineering. For decades, cybersecurity training has taught users to look for "red flags" such as poor grammar, spelling errors, and awkward phrasing. AI has effectively scrubbed these signals from the attacker’s toolkit.

Robert Knake, a panelist and partner at Paladin Capital who previously served as the first deputy national cyber director for strategy and budget at the White House, observed that the barriers to entry for high-level deception have vanished. "A year ago, we still had email messages in our inbox that had misspellings that were not colloquial English, that were easy to identify if you were vigilant. Now, all those signals are gone," Knake stated.

Beyond text, agentic AI can now coordinate "vishing" (voice phishing) and deepfake video attacks with terrifying precision. A notable case involved a multinational firm in Hong Kong where an employee was tricked into transferring $25 million after attending a video conference with what appeared to be the company’s CFO and other staff members—all of whom were AI-generated deepfakes. Agentic AI takes this a step further by allowing these deepfakes to interact in real-time, responding to questions and navigating social cues autonomously.

The Regulatory Debate: Liability and the "Safe Harbor" Concept

The consensus among the Berkman Klein Center panelists was that the current "wild west" of AI development cannot persist if national security is to be maintained. However, the path to regulation is fraught with technical and economic hurdles. Robert Knake argued that the federal government must shift the burden of security from the end-user to the software manufacturers.

Knake proposed a regulatory "Safe Harbor" framework to balance innovation with accountability. Under this model, software companies would not be held liable for every unforeseen bug, as a "zero-error" requirement would effectively stifle the software industry. Instead, companies would be granted legal protection only if they adhere to established best practices, such as maintaining an updated inventory of open-source packages and utilizing known secure versions of code.

"We’re not at a place where we can say any error in your software that leads to a harm, you need to be responsible for. That will kill off software development," Knake explained. "But we could create a safe harbor… If you haven’t done [the basics], you should be [liable]."

This approach aims to address the systemic issue of "technical debt" and insecure coding practices that have long plagued the industry. By attaching financial and legal consequences to negligence, regulators hope to force a "secure-by-design" mentality into the AI development lifecycle.

Technical Barriers to Governance and Oversight

Despite the appetite for regulation, James Mickens cautioned that implementing such schemes is technically complex. He noted that the threat model for AI is fundamentally different from traditional software. For decades, giants like Microsoft and Amazon have used internal stopgaps to prevent traditional breaches. However, agentic AI introduces the problem of "adversarial prompting" and "command injection," where a human user—or another AI—can trick a model into bypassing its own security protocols.

"Essentially, there’s some human in a chair that’s outside of the data center who’s sending evil commands to the code that’s running in the data center and otherwise trying to trick it into being evil with AI," Mickens said. This fluidity makes it difficult to define what constitutes a "compliant" or "secure" AI system in a way that can be codified into law.

Josephine Wolff, associate dean for research and professor of cybersecurity policy at the Fletcher School at Tufts University, added that the sheer scale of modern networks makes proactive vulnerability management a Herculean task. She questioned whether corporations are even capable of maintaining the "documentation and inventories" required for effective regulation. "Can you inventory all of the code that’s running on your computers so that if there’s a vulnerability… you can at least know where you need to look?" she asked.

The Perils of "Hacking Back" and Digital Vigilantism

As cyberattacks become more frequent and damaging, some voices in the private sector have advocated for the right to "hack back"—to engage in offensive counter-operations against attackers. The panel of experts, however, was nearly unanimous in their opposition to this trend.

Wolff argued that deputizing the private sector to engage in cyber-retaliation would lead to "greater chaos." She warned that while large firms like Google might have the resources for surgical strikes, smaller, less-regulated firms might launch ill-conceived attacks against foreign infrastructure, potentially triggering international conflicts. "I think you would have a whole bunch of much crazier firms with many fewer lawyers feeling like, here’s our opportunity to take on North Korea. And that doesn’t seem to me like a safer world," Wolff said.

Mickens provided a chilling vision of a world where agentic firewalls autonomously retaliate against perceived threats. He compared this to high-frequency trading in the financial markets, where algorithms react to one another in milliseconds. In a cybersecurity context, this could lead to a rapid, automated escalation where algorithms trade "offensive maneuvers" back and forth, potentially crippling global internet infrastructure before a human can intervene. "I don’t think we want to get into that world for the same reason that… we don’t want to sort of deputize vigilantes in the physical world," Mickens added.

The Future of Identity: Verification vs. Privacy

To combat AI-driven impersonation and fraud, the experts suggested that the digital world may need to move toward a more robust system of identity verification. Knake argued that in a world of deepfakes, "knowing with certainty who we are dealing with" is the only way to restore trust in digital interactions.

However, this proposal raises significant privacy concerns. Mickens noted that many vulnerable populations—such as victims of domestic abuse, political dissidents, or whistleblowers—rely on pseudonymity for their safety. "One reason digital IDs have traditionally struggled is that there are many scenarios in which someone wants to be identified as part of their identity, but not the full identity," Mickens said. He emphasized that any move toward universal digital identification must solve these "practical problems" to avoid creating a surveillance state that endangers the very people it seeks to protect.

Conclusion: Harnessing the "Guardian Angel" AI

While the risks of agentic AI are profound, the experts concluded with a note of cautious optimism. The same technology that empowers hackers can also serve as a "guardian angel" for users. Knake envisioned a future where agentic AI sits "over your shoulder" on every device, monitoring for signs of a "kill chain" in fraudulent schemes and intervening in real-time to protect the user.

The transition to a secure AI-integrated world will require a tripartite effort: rigorous government regulation to establish liability, a commitment from the private sector to prioritize security over speed, and a global dialogue on the ethical boundaries of autonomous systems. As the panel at the Berkman Klein Center made clear, the technology is already here; the challenge now lies in finding the "market players" and political will to build the defenses necessary to control it.