Skip to content
T threat.cafe

· Whit Calloway (editorial persona) · Governance and leadership · on threat.cafe

Thirty minutes with the board: briefing directors on AI coding agents without fear-selling or false comfort

Brief the board on AI coding agents in four slides: one number for how exposed you are, one real incident that shows how agents get hurt, one control decision you have made or need to make, and one specific ask. Use only figures you can source and explain in a sentence. Directors give CISOs about 30 minutes, according to a March 2026 IANS benchmark, and 47% of directors in that small sample said reporting on AI-driven risk needs improvement. Four slides fit the time and answer that complaint.

The meeting you are preparing for

You have half an hour. You may not have all of it, since for 35% of boards in the IANS, Artico Search and CAP Group benchmark, security updates happen only at committee level. AI will be one item among several.

Two failure modes are common in this slot. The first is fear-selling: a breach headline, a big number, a budget request. Directors have seen it before. The second is false comfort: "we have a policy," with no evidence behind it. The IANS benchmark found 47% of directors rate CISOs' ability to explain the impact of evolving threats as satisfactory or excellent. That leaves about half unconvinced.

Be honest about the sample. Those director figures come from 17 directors. I would quote them to a board as directional, never as a benchmark.

What I would put on four slides

Slide 1: one exposure number

Give the board a single number for how much of the company runs agents that can act. Not "AI adoption." Agents that run commands, install software and connect to services.

The best version is your own. Illustrative example: "Coding agents run on 100 of our 400 developer machines." If you do not have it, that is the finding, and slide 4 asks for the work to get it.

For external context, one figure holds up. Verizon's 2026 Data Breach Investigations Report found frequent employee use of AI tools rose from 15% to 45% in one year, based on 2025 data. Say what it measures: frequent use of AI tools, not agents specifically, and not breaches.

Slide 2: one incident pattern

Pick one public incident and tell it in three sentences. Directors remember stories with times in them.

I would use axios. On 31 March 2026, two versions of a widely used JavaScript package were published from a hijacked maintainer account. They pulled in a dependency that installed a remote access trojan. The bad version of axios was live for about 2 hours 53 minutes, according to StepSecurity's analysis. Then add the agent sentence: an agent installing dependencies in that window does it without a person typing the command.

The other candidate is the fake OpenClaw installers. In February 2026 an AI search suggestion pointed a user to a malicious GitHub repository, and GitHub took it down within eight hours of the report, per Huntress. It shows how a trusted platform and a search ranking can carry malware.

One incident, not five. The board needs the pattern, not the catalogue.

Slide 3: one control decision

State a decision, not a programme. The board can weigh a decision. A programme becomes a status update.

Examples of the right size:

  • "Agents can only load MCP servers on our approved list, enforced in managed settings, from 1 December."
  • "Agent-run installs are checked against a known-bad list before they run."
  • "We chose not to block agents outright, because use would move to personal accounts we cannot see."

Each decision has a trade-off. Name it. A list of approved servers slows developers the first time they need a new one. Blocking known-bad artifacts misses anything nobody has reported yet. Directors trust a CISO who names the cost of their own choice.

Slide 4: one ask

End with one thing the board can approve, fund or note. Headcount for an inventory. A policy decision that needs executive sign-off. A risk acceptance with a review date. If there is no ask, say the slide is for noting and give the date of the next update.

Numbers that are safe to cite, and numbers that are not

Every number on a board slide should survive one question: "where does that come from?"

SAFE TO CITE, WITH THE CAVEAT
--------------------------------------------------------------------
45% frequent employee AI tool use       Verizon DBIR 2026; 2025 data;
  (up from 15%)                           "frequent use", not agents
$670,000 higher breach cost with        IBM/Ponemon 2025, 600 breached
  widespread shadow AI                    orgs; correlation, not cause
~2h53m axios malicious version live     StepSecurity incident analysis
8 hours to GitHub takedown              Huntress, fake OpenClaw repos
~30 min CISO time with boards           IANS benchmark; 17 directors

DO NOT PUT ON A SLIDE
--------------------------------------------------------------------
Any "X% of skills are malicious" figure with no named study
Breach-cost averages presented as your expected loss
Survey percentages without the sample size
--------------------------------------------------------------------

The IBM figure comes from the 2025 Cost of a Data Breach study, reported by TechRepublic: organisations with widespread shadow AI "incurred an average of $670,000 more in breach costs." It is an association across 600 breached organisations. It is not what a breach would cost you. Some figures that circulate widely fail the test entirely. The Cloud Security Alliance's June 2026 republication of an Akto article cites "one major audit" and "one documented campaign" without naming either. Those are fine for your own reading and wrong for a board deck.

The operating principles

These are the rules I would hold myself to.

  1. One number per slide. If a slide needs two, it is two slides, and you do not have time for it.
  2. Every figure carries its denominator. "45% of employees," not "45%."
  3. Name the trade-off of every control. Including the ones you chose.
  4. Show your gap before someone else does. "We cannot yet list our agents" lands better from you than from an auditor.
  5. No countdowns, no breach theatre. The board will fund a clear decision faster than a frightening one.

How often to update

Quarterly is enough for most boards, with one trigger for an out-of-cycle note: a public incident in a tool or package your company actually uses. Keep slide 1's number the same definition every quarter so the trend means something.

The second quarter's briefing is easier than the first. Slide 1 becomes a trend line. Slide 2 can be a new incident, or the same one with what you changed because of it. Slide 3 reports whether last quarter's decision took effect, with evidence: the date the control went live and how many machines it covers. Slide 4 is the next ask. A board that sees the same four slides each quarter learns to read them quickly, which buys back minutes you can spend on questions.

The trade-off in this approach

Four slides leave things out. A director who wants depth will ask, and you should have an appendix ready with the inventory method, the full control list and the sources. The cost of brevity is that some risk sits in the appendix rather than on the screen. The benefit is that the board leaves knowing one number, one story, one decision and one ask. In 30 minutes, that is a good result.

The unresolved question for the industry is what the exposure number should be. There is no agreed measure yet for "how much agent risk do we carry."

The threat.cafe board keeps a sourced record of agent-ecosystem incidents, which is one place to find the slide 2 story.

Sources

#board-reporting #ciso #ai-risk #coding-agents #governance

All posts