The new security agenda was written for enterprises. Your company inherits the threats, not the budget.

On 28 October, a16z published The New Security Agenda. It opens with a line worth reading twice: "For years, the industry prepared for quantum computing to rewrite the rules of cybersecurity. Then AI got there first."

It is a good essay, and it was written for somebody else. The authors surveyed 100 CISOs at companies with more than 1,000 employees. It says nothing about small and mid-size firms. If you run a company of 20 to 200 people, you will not buy an autonomous penetration-testing platform this quarter.

You will, however, face the same attackers, at the same speed, through the same AI tools your team installed last month. That is the argument of this piece: you inherit the threats, not the budget. What follows is what the agenda says, what the latest data adds, and which parts of it a company with no security team can actually use.

What the agenda says

The survey results are the useful part, because they show where large buyers think the risk has moved:

• Spend is rising. 87% of CISOs expect security spending to increase in 2027, driven by AI-enabled threats and the attack surface their own AI deployments create.
• AI touches the whole priority list. All seven of the most selected 2027 initiatives involve AI. The biggest gains since last year are internal AI agent security, automating the security operations centre, and offensive security.
• Agents are already inside. 92% report at least limited use of employee generative AI tools, 83% use coding assistants or agents, and 55% have MCP servers or other agent integrations in production.
• Security is slowing adoption. 63% say it holds back at least some AI projects.
• The old tools are blind to agents. An agent runs with the employee's own permissions, so "to the tools built for the last decade of endpoint security, an agent reading a file or calling an API looks exactly like the employee doing it."
• Finding is easy, fixing is slow. In the words of one fintech CISO: "you're able to find vulnerabilities faster, but you're not able to patch faster."

The essay's commercial conclusion is that incumbents get the first call, and startups win where there is a capability gap. For a large company that is a buying guide. For you it is a weather report.

The weather: what the latest data says

Four findings from the past six months matter more than the rest.

1. Exploitation is now the front door. Verizon's 2026 DBIR found vulnerability exploitation was the initial access route in 31% of breaches, the first time in 19 years it has passed stolen credentials. Verizon says AI is shrinking the window for defenders "from months to mere hours".

2. The window has closed, and sometimes inverted. Mandiant's M-Trends 2026 estimates mean time to exploit at about minus seven days: attackers often move before a patch exists. It also found the handoff from initial access to a second criminal group fell from over eight hours in 2022 to 22 seconds in 2025. CrowdStrike puts average breakout time at 29 minutes.

3. Offence got cheap. Anthropic's N-day study found its restricted Mythos Preview model built a working Firefox exploit within an hour of a patch release, and Windows kernel exploit chains at roughly $2,000 each. That model is not publicly available, so read it as capability, not as confirmed criminal use. But the direction is plain.

4. AI attacks already cost more. IBM's 2026 Cost of a Data Breach report says one in four malicious breaches was AI-enabled, averaging about $6 million against a $4.99 million global average. Voice phishing is also rising: M-Trends has it at 11% of initial access, ahead of email phishing at 6%.

None of those figures is filtered by company size. Attackers automate, and automation does not check your headcount.

Where SMEs are actually exposed

Here we part ways with the enterprise agenda, because the problems are different in kind.

You are already being hit. The UK government's Cyber Security Breaches Survey 2025/26 found 42% of micro, 46% of small and 65% of medium businesses had a breach or attack in the past year.

You recover worse. Sophos's State of Ransomware 2026, covering 17 countries including India, the UK, Singapore and Australia, found only 34% of organisations with 100 to 250 staff stopped an attack before data was encrypted or held to ransom, against 46% at 3,001 to 5,000 staff. 79% of attacks began with compromised identities, which is a password-and-MFA problem, not an AI problem.

Your AI governance is thinner than your AI use. The same UK survey found only 24% of AI-using firms have cyber practices specific to AI risk, and only 15% of businesses review the risk of their immediate suppliers. Gartner's September survey of 297 security leaders reports that 54% of organisations lack a defined way to restrict AI agent access. Enterprises are not far ahead of you here, which is the good news.

You have no review capacity. The agenda's central tension is that AI finds problems faster than humans fix them. In a 40-person company there may be no named owner for patching at all. The backlog does not just grow, it sits on the internet while exploitation starts before the patch.

Your customers will push their risk onto you. Large buyers cannot afford to be breached through a supplier, so security questionnaires and contract clauses flow down. Regulation adds pressure rather than relief: the EU Digital AI Omnibus pushed high-risk AI obligations back to December 2027, but NIS2 and the Cyber Resilience Act still reach small suppliers through their customers. Check each against your own situation, because the dates are still moving and several of our sources were secondary.

Your insurer is changing the terms. Cyber cover increasingly assumes multi-factor authentication, endpoint detection and tested backups before it will quote, and some general-liability policies now carry optional generative-AI exclusions. Ask your broker what yours says about agents and deepfake fraud.

What transfers, and what does not

The agenda says	A 5,000-person company does	A 50-person company should do
Secure the AI employees already use	Buys an agent-visibility platform	Keeps a one-page register of every AI tool and agent, with owner and what it can touch
Test continuously	Runs autonomous offensive security	Scans internet-facing systems monthly and retests after any change
Fix faster than you find	Automates remediation tickets	Names one owner, sets a patch deadline in days, and keeps a short list of systems that cannot wait
Add AI to the SOC	Deploys SOC agents	Buys managed detection and response, and checks it covers identity alerts
Defend against human risk	Runs deepfake simulations	Agrees a call-back rule for money and credential requests, and tells staff plainly why
Choose among noisy vendors	Runs proofs of concept on seven tools	Buys fewer tools, from platforms you already pay

The last row is the most useful. The essay notes that 44% of CISOs default to their existing security platform and that one CISO refused to run seven proofs of concept. If a team with a budget finds the market too noisy, a company without one should not try to out-shop it.

Three bets that hold up

Bet 1: shrink the blast radius, because you cannot stop the attempt. The agent problem is the same one we cover in You cannot stop prompt injection. You can cap what it costs. An agent that can read private data, see outside content and send mail out is a leak waiting for a sentence. Remove one of the three. This is the cheapest control in the whole agenda.

Bet 2: make patching boring and fast. With exploitation now the top entry route, the old monthly cycle is a liability. Internet-facing systems and anything on a known-exploited list get days, not weeks. Everything else can wait. Speed on a short list beats thoroughness on a long one.

Bet 3: verify people, not faces. Mobile-centric social engineering succeeds around 40% more often than email phishing, per Verizon, and a convincing cloned voice no longer takes a specialist. The defence is a rule, not a gadget: no payment, credential or access change is approved on the strength of a call or a video. You call back on a number you already hold. The widely reported Arup deepfake case, in which a finance worker approved about US$25 million in transfers on a video call full of fakes, is the argument for writing that rule down.

What not to say to your board

Two claims circulate that you should drop. One is that 60% of small firms close within six months of an attack: that is a myth, and the National Cyber Security Alliance has disowned it. The other is that 43% of attacks target small businesses, which comes from an old Verizon figure that is no longer current. The real picture needs no embellishment: roughly one in five small firms say an attack threatened their viability, and the cost lands on cash and attention you cannot spare.

Be honest about the data's limits too. Several figures above rest on vendor surveys, and the a16z survey measures intentions and adoption, not outcomes.

Your next 30 days

1. Week 1: list. Every AI tool and agent, who owns it, what it can read, what it can send or change. Include the browser extensions and meeting bots your team added without asking. Our one-page AI guidelines give you a template.
2. Week 2: lock identity. MFA on email, finance, admin consoles and every AI tool. Remove stale accounts. Most attacks that reach encryption start with a compromised identity.
3. Week 3: set the patch rule. One owner. Internet-facing systems patched within a week, critical ones sooner. Write the exceptions down.
4. Week 4: write the call-back rule and cap the spend. Brief everyone who can move money. Then set limits on what agents can spend, as in what a sane agent spend cap looks like.

At the end of the month, ask your insurer and your largest customer what they would want to see. You will be surprised how much of it you now have.

Working with us

This is the territory of our six-step AI security plan, and a fractional CAIO can own the register, the permissions and the supplier conversations. If you would rather measure first, the free scorecard takes three minutes.

The enterprise agenda tells you where the attackers are heading. You do not need their budget to move out of the way.