The AI Cybersecurity Poverty Line

The line was drawn fifteen years ago

In 2011, Wendy Nather, then a research director at 451 Research, described what she called the security poverty line: the point below which an organization cannot afford an effective level of security, let alone compliance, mostly because its IT budget is too thin to begin with. The idea stuck. Nather and Andy Ellis expanded it at RSA Conference in 2013 (SecurityWeek), Cisco's Jeetu Patel told RSAC 2022 it would eventually become a human rights issue (The Register), and an RSAC 2025 panel revisited it fourteen years on.

Nather's current framing is the useful one: the line below which an organization cannot be effectively secured. Below it, security is not a program. It is borrowing, begging and bartering (Infosecurity Magazine).

From where a security practice sits, working with companies on both sides of that line, the pattern in 2026 is consistent. AI has not moved the line. It has moved the ground under it, and for most organizations it has moved it the wrong way.

Offense now runs at machine speed

The evidence that attackers have automated the hard parts is no longer anecdotal. In November 2025, Anthropic disrupted a state-sponsored espionage campaign that used its coding agent as an autonomous attack framework against roughly thirty targets. AI performed an estimated 80 to 90 percent of the work, with human intervention needed at perhaps four to six decision points per campaign (Anthropic). Google's Threat Intelligence Group reported in May 2026 the first AI-developed zero-day exploit observed in the wild, and agentic workflows making tactical decisions at machine speed. CrowdStrike's 2026 Global Threat Report puts average eCrime breakout time at 29 minutes, with the fastest observed at 27 seconds.

The front door has changed too. Verizon's 2026 Data Breach Investigations Report found exploitation of vulnerabilities to be the top initial access vector, at 31 percent of breaches, and attributes part of that shift to AI collapsing the time from disclosure to exploit. On the human side, Microsoft's Digital Defense Report 2025 measured a 54 percent click-through rate for AI-automated phishing, against about 12 percent for conventional phishing.

Governments read it the same way. The UK National Cyber Security Centre judged in 2025 that AI will almost certainly make intrusion operations more effective and efficient, and that there will almost certainly be a digital divide between systems keeping pace with AI-enabled threats and a large proportion that are more vulnerable (NCSC). In June 2026 the Five Eyes agencies jointly warned that AI is shrinking the discovery-to-exploitation window.

The data is not uniform. Google noted as recently as February 2026 that it had not yet seen fully agentic attacks in the wild, and Verizon found that AI-written phishing doubled in volume without a measurable rise in success. The direction, though, is one way. No organization hires its way to a 29-minute response window. The defense that keeps pace is one that also runs at machine speed.

Machine-speed defense has a price tag

The AI that defenders need is metered, tiered and priced for organizations that already have budget. Prices below are as listed in October 2026.

Start with the models. Every major provider sells access in rungs, and the rungs are defined by exactly what a security team cares about: how much it can use, how much context it can load, and which models it can reach. Anthropic offers a free tier, Pro at $20 a month, and Max from $100 a month with five or twenty times Pro's usage, with Team and Enterprise priced separately (Claude pricing). OpenAI's free tier is limited to lighter models and a smaller context window than its paid tiers. Google puts Deep Think and its most capable agent features in AI Ultra, at $99.99 or $199.99 a month (Google). At the API layer, where real automation lives, frontier models are billed per million tokens, and an agent that reads every alert, log line and ticket consumes them by the hundreds of millions.

The security platforms built on those models follow the same shape. Microsoft Security Copilot is sold in provisioned security compute units at $4 per unit per hour, roughly $2,900 a month for one always-on unit before anyone has asked it a question; Microsoft now includes Security Copilot agents with Microsoft 365 E5, which is to say the included capability scales with how large a customer already is (Microsoft). CrowdStrike's self-serve Falcon Go is capped at 100 devices, and Charlotte AI sits in the quoted tiers (CrowdStrike). Enterprise threat intelligence platforms list in the tens to hundreds of thousands of dollars a year on public government price schedules.

None of this is unreasonable for what it does, and that is the point. Machine-speed defense is a real product, and real products cost real money. The capability to meet AI-driven offense is concentrating where capability has always concentrated: large enterprises with large licenses. A 40-person company does not have a provisioned compute unit. It has a shared login and a rate limit.

The gap is widening

If the poverty line were holding steady, small and under-resourced organizations would be roughly where they were five years ago. They are not.

They are the targets. Verizon's 2026 data puts small organizations at 96 percent of ransomware victims, and its companion breach impact study found the worst 2.5 percent of small-business breaches cost more than 7 percent of annual revenue. Sophos's State of Ransomware 2026 found that 34 percent of organizations with 100 to 250 employees stopped an attack before encryption, against 46 percent of those with 3,001 to 5,000 employees, and that median recovery costs reached $1.7 million. CISA has a name for the sectors that live here, hospitals, K-12 schools and water utilities: target rich, cyber poor.

They are not adopting AI defense. CrowdStrike's 2025 State of SMB Cybersecurity report found that only 11 percent of small and mid-sized businesses use AI-powered defenses, that more than half of companies under 50 employees spend less than 1 percent of their budget on security, and that two-thirds cite cost as the barrier to upgrading. Sage and IDC found in May 2026 that 84 percent of micro businesses are unprepared for AI threats. Only 23 percent of US small businesses have a documented AI policy.

The people who would run the defense are under pressure. The ISC2 2025 Cybersecurity Workforce Study, drawing on 16,029 respondents, found 36 percent reporting budget cuts, 24 percent reporting layoffs, and 59 percent citing critical or significant skills needs, up from 44 percent the year before. Only 28 percent of respondents had already integrated AI tools into their operations. Seventy-two percent agreed that reducing security staff significantly increases the risk of a breach (ISC2).

Put the two halves together. Offense is automating and accelerating, measured in minutes. Defense at that speed is sold in compute units, premium seats and six-figure contracts. The organizations most likely to be hit are the least likely to afford the countermeasure and the most likely to be losing the people who would deploy it. That is not a static poverty line. It is a line moving away from the organizations standing below it.

What works below the line

Nobody has fixed the economics yet. Nather has floated a pro bono cyber peace corps, and the RSAC 2025 panel spent most of its time on policy remedies, both worth pursuing. In the meantime, these are the moves that hold up in practice for organizations that cannot buy their way out.

1. Fix the basics first, because that is where machine-speed attacks land. Exploitation of known vulnerabilities is now the leading way in. A patched edge device and enforced MFA do not care how fast an attacker's agent runs. Most ransomware victims are not beaten by a zero-day.
2. Pool, do not buy. The clearest lesson in the pricing pages is that AI defense is cheap per unit and expensive per organization. Spreading one compute commitment, one premium seat or one detection platform across many organizations, the model managed security providers and virtual CISO practices already run, is how the math works below the line.
3. Use what is already paid for. Included Security Copilot capability, bundled EDR features and the AI already inside the email gateway often go unused. Inventory current licenses before buying a new platform.
4. Write the AI policy before it is needed. Most small companies have neither an AI security policy nor an AI use policy. That is a document, not a product. Our AI governance checklist is a starting point.
5. Measure time to respond, not tool count. If breakout is measured in minutes, the question that matters is how long it takes to notice and contain. If the honest answer is days, no subscription changes that until the process exists.

The line is real

Fifteen years after Nather drew it, the security poverty line is as relevant as it has ever been. What has changed is what sits on the other side of it. It is no longer just a bigger budget or a bigger team. It is machines defending against machines, and the entry fee is rising. Organizations that can afford the best AI access will defend at the speed of the attack. Organizations on free tiers, rationing context and waiting out rate limits, will not. Whether the industry accepts that outcome is a choice, not an inevitability.

For the underlying numbers on smaller organizations, see our SMB cybersecurity statistics for 2026 and the AI cybersecurity incident report. For building a defensible budget below the line, see the cybersecurity budget template.

Sources

• Wendy Nather, "The Security Poverty Line, and junk food," December 23, 2011.
• SecurityWeek, "Software Development Below the Security Poverty Line," May 25, 2016.
• Infosecurity Magazine, "Profile Interview: Wendy Nather," February 18, 2020.
• The Register, Cisco at RSA Conference, June 6, 2022.
• RSA Conference, "Bridging the Cyber Poverty Gap," May 1, 2025.
• Anthropic, "Disrupting the first reported AI-orchestrated cyber espionage campaign," November 13, 2025.
• Google Threat Intelligence Group, adversarial use of AI reports, February 12 and May 11, 2026.
• CrowdStrike, 2026 Global Threat Report, February 2026; State of SMB Cybersecurity, May 2025.
• Verizon, 2026 Data Breach Investigations Report and Breach Impact Study, 2026.
• Microsoft, Digital Defense Report 2025.
• UK National Cyber Security Centre, "Impact of AI on cyber threat from now to 2027," 2025.
• Five Eyes agencies, joint statement on AI and cyber risk, June 22, 2026.
• Anthropic, OpenAI, Google, Microsoft and CrowdStrike published pricing, October 2026.
• Sophos, State of Ransomware 2026.
• CISA, Target Rich, Cyber Poor.
• Sage and IDC, May 2026; Pax8, July 2026.
• ISC2, 2025 Cybersecurity Workforce Study, December 2025.