How to Restrict Application Network Access via System Firewalls
Network Visibility: The True Antidote to AI-Driven Cyber Threats
Last month, a formidable coalition of tech titans—including OpenAI, Anthropic, Microsoft, Cisco, and CrowdStrike—released an open letter urging a unified front on cybersecurity. Initially, I braced myself for a thinly veiled sales pitch disguised as a public service announcement. However, I walked away largely agreeing with their core thesis: artificial intelligence is simultaneously supercharging both offensive and defensive cyber capabilities, and we have a rapidly closing window to ensure the good guys maintain the upper hand.
That said, I believe the letter buries its most vital insight beneath layers of industry rhetoric.
Old Vulnerabilities at a Terrifying New Velocity
When the letter critiques the inadequacy of our current security posture, it lists the usual suspects: excessive user permissions, misconfigurations, fragile authentication, unpatched software, and the suffocating technical debt of legacy systems. I’ve seen these exact same flaws cited in every post-mortem report over the past fifteen years. None of this is AI’s fault.
What AI does change is the sheer velocity at which these pre-existing vulnerabilities are discovered and weaponized. The statistics speak for themselves:
- The Speed of Creation: According to IBM X-Force’s 2026 research, drafting a targeted phishing email used to take a human hacker roughly 16 hours. Generative AI has slashed that time to just 5 minutes—yielding a 200-fold increase in attacker productivity.
- The Speed of Victimization: Verizon’s 2025 Data Breach Investigations Report (DBIR) notes that the median time it takes for a victim to click a malicious link is a mere 21 seconds.
The attack vectors haven’t changed, but the speed has. Human-led detection processes simply cannot intervene within a 21-second window.
The Critical Infrastructure Blind Spot
This acceleration is most alarming in sectors where the stakes are highest, such as healthcare and municipal water treatment—environments heavily reliant on legacy hardware untouched since the 2010s. Consider Forescout’s latest scan, which identified approximately 3.4 million exposed remote access servers globally. Alarmingly, around 670 of those were VNC connections granting direct, unauthenticated access to critical OT and ICS control panels. Furthermore, over 40% of the exposed RDP servers were running Windows 10, an OS that reached its end-of-support life last October.
You cannot simply parachute an AI-powered SOC copilot into a network running 2011 PLC firmware and declare the network secure. Flashy demos don’t patch legacy vulnerabilities.
The letter hits the nail on the head regarding one crucial reality: when critical systems cannot be taken offline for patching, you must rely on verifiable compensating controls.
Historically, this is exactly the role Network Access Control (NAC) was designed to play—detecting a device the millisecond it connects, assessing its authorization, and isolating it without needing to alter the endpoint itself. Unfortunately, this security discipline has been chronically underfunded. That financial neglect is arguably a far larger contributor to our current security woes than the open letter acknowledges.
The Machine Identity Explosion
The most actionable mandate in the entire open letter is buried in a section targeting AI developers: the urgent need to “ensure agentic identities are traceable and accountable.” This aligns perfectly with arguments I’ve been making all year, heavily inspired by CoSAI’s framework for agentic identity.
The scope of this challenge is staggering. While estimates vary, Palo Alto Networks’ 2026 Identity Security Landscape report indicates an average enterprise now harbors 109 machine identities for every single human identity—a massive jump from the 82-to-1 ratio seen just a year prior. Autonomous AI agents are driving the bulk of this surge.
The security implications are profound. The 2026 Infrastructure Identity Survey revealed that networks enforcing strict least-privilege access for AI agents experienced only a 17% incident rate, compared to a devastating 76% for systems allowing over-privileged agents. There is no clearer proof that default access must be ruthlessly restricted.
Back to the Fundamentals
It is refreshing to see accountability and traceability highlighted by major tech firms, but drafting a letter doesn’t magically engineer the solution. In practice, securing AI means treating every autonomous agent like an untrusted device joining your network. You must identify it before it executes a task, strictly limit its scope, meticulously log its actions, and possess the capability to sever its access instantly.
We need to stop treating AI agents like neglected service accounts and start managing them as active entities within our Identity and Access Management (IAM) stacks. They must be auditable, controllable, and expendable the moment they exhibit anomalous behavior. The industry learned this lesson the hard way with legacy machine credentials; the threat has simply multiplied in scale and autonomy.
Ultimately, the coalition’s open letter isn’t wrong, and achieving this level of consensus among tech rivals is commendable. But piling more AI security tools on top of an infrastructure that cannot even identify what is plugged into it is a recipe for failure. You cannot rely on an identity program designed for a 1:1 human-to-machine ratio in a world where that ratio is 1:100.
The core tenets of cybersecurity remain unchanged: Know exactly what is on your network, understand what it has access to, and ensure you have the kill-switch to revoke that access instantly. AI will undoubtedly help us identify network gaps faster, but it is entirely useless if the foundational access-control layer isn’t there to actually enforce the fix.