Dave Merkel

Attackers already use AI, so defenders will have to as well or lose ground. Figuring out how becomes part of the job, not a separate project.

The post names browser-embedded AI, sensitive data entered into tools such as OpenAI or Gemini, and prompt injection that triggers odd model behavior as places attackers will push, using employees to reach data they cannot access directly.

Read original

They call that backwards: speed without accuracy is reckless, and accuracy without speed is useless.

• AI should not get the wheel on day one; it should earn autonomy through demonstrated outcomes.
• Humans should watch AI, and AI should watch humans—a necessary part of the architecture, not a limitation to engineer away.

The post sums up the argument in a few sentences and links to a published framework for deciding where AI belongs in a security operations center; the supporting detail is behind the link.

Read original

Real security, they argue, requires transparency, independent auditing, accountability, and diversity of approach rather than an opaque monoculture.

• The post warns Western democracies not to treat India's policy as a roadmap, noting that corporations have already surrendered data access under regulatory pressure and that voluntary frameworks, limited access, and oversight mechanisms can expand into mandates and rubber stamps.
• The target is centralized, opaque, judicially unsupervised software—not regulation as such.

India's Sanchar Saathi pre-installation order and its rapid reversal anchor the post.

Read original

The piece presents talking points for practitioners, with a short reason for each call rather than a data-backed forecast, and closes by asking other security leaders what they are watching.

The post illustrates a widening attack surface with AI in the browser, sensitive data entered into tools such as OpenAI or Gemini directly or through employees, prompt injection producing strange model behavior, and the prospect of something unimagined also emerging.

Read original

It answers “what does this person actually forecast, and how do they reason about it?” before you read anything narrower.

Merkel lays out five predictions for 2026.

• Federal cyber policy heads toward disorder, with public-facing capabilities like CISA dismantled while the administration takes a conciliatory stance toward North Korea, China, and Russia.
• M&A should pick up, especially around AI-enabled security companies and distressed assets, while IPOs stay slow.
• AI will not remove cybersecurity jobs, they argue, because the adversary is still a human being and defenders still need to spot patterns, adapt, and think like attackers.
• But AI will change those jobs, because attackers are already using it.
• The attack surface keeps widening: AI in the browser, sensitive data inside tools like OpenAI or Gemini, prompt injection producing odd model behavior, and something nobody has imagined yet.
• The fifth prediction is simply that the year will contain a surprise.
• Each point comes with a short reason rather than a data table, and the post closes by asking other security leaders what they are watching.

For brands planning a campaign

Useful for a keynote or advisory framing around forward-looking themes.

• The five-part structure is ready-made for an event agenda or fireside format, and the AI-and-jobs argument offers a defensible counterpoint to “AI replaces analysts” messaging many vendor audiences have already heard.
• The predictions are assertions, not evidence-backed forecasts, so use them as talking points rather than research findings to cite.

Open it to see whether they can take a public position and defend it.

India ordered new smartphones to ship with a state-run cybersecurity app, then reversed the order.

• Merkel's response is not relief but warning.
• Mandating government-controlled security software on every device, they argue, creates a single point of compromise: one breach hands attackers or the government direct access to hundreds of millions of devices.
• In this account, real security requires transparency, independent auditing, and accountability—all absent when software is opaque and has no judicial oversight.
• Diversity of approach beats monoculture.
• The second half turns to Western democracies: corporations have already capitulated to government pressure for data access rather than face regulatory consequences.
• Any “voluntary” framework becomes mandatory, any “limited” access expands, and any oversight becomes rubber-stamping.
• The post ends with the argument that governments create new threats rather than address existing ones.

For brands planning a campaign

Relevant when a campaign needs a credible voice on privacy, regulation, or government overreach—areas that sit awkwardly beside typical product messaging.

• The argument is opinionated and clearly reasoned, making it a fit for panel moderation, policy commentary, or an advisory role where a client wants pushback rather than agreement.
• The position criticizes government mandates generally; campaigns tied to public-sector or compliance-driven products should weigh that stance before approaching.

It is short, but states a position and points to a longer framework.

Merkel argues that most AI implementations optimize for speed or accuracy at the expense of the other, and that in SecOps this is backwards: speed without accuracy is reckless, accuracy without speed is useless.

• Their central idea is that AI earns autonomy through demonstrated outcomes rather than being handed control on day one.
• Humans should watch AI, and AI should watch humans—this is how the architecture needs to work, not a temporary limitation.
• The post closes by pointing to a published framework for deciding where AI belongs in a SOC, with hashtags for MDR, cybersecurity, and security leadership.
• The reasoning is compressed into a few sentences; the supporting detail sits behind the linked framework rather than in the post itself.

For brands planning a campaign

A natural anchor for a keynote or advisory engagement on AI in security operations.

• The mutual-oversight line and the earned-autonomy idea give a speaker concrete, quotable positions to build a talk around.
• The linked framework carries the technical detail; if a campaign depends on that depth, confirm the material is available and current.

The post closes by asking other cybersecurity leaders what they are watching for in 2026.

A numbered list of year-ahead predictions for practitioners, mixing macro policy and market calls with operational security arguments and an open question to peers.

For brands planning a campaign

Fits a keynote or advisory briefing on the 2026 threat and regulatory landscape, or a moderated discussion. The structured predictions give a speaker clear segments and a natural hook for audience participation.

Examples: Source 1

The post describes a single point of compromise that, if breached, hands attackers or the government direct access to hundreds of millions of devices.

• They tie real security to transparency, independent auditing, accountability, and diversity of approach over monoculture, noting that centralized, opaque software with no judicial oversight offers none of those.
• The post warns that Western democracies facing similar pressure have already shown a pattern of capitulating to government data-access demands, so any voluntary framework, limited access, or oversight mechanism is likely to become mandatory, broader, and a rubber stamp.
• The recommended path is privacy-respecting security built on transparent standards, independent verification, and distributed architecture, treating India's attempted policy as a cautionary tale rather than a roadmap.

The post argues against government-mandated, centrally controlled security software—not all regulation. Its stated failure mode is a single point of compromise combined with missing transparency and oversight, not that government systems are always compromised.

For brands planning a campaign

Useful for a policy panel, a privacy-and-trust webinar, or a vendor positioning conversation about transparent, independently verified, distributed security. The post makes a clear, quotable argument against centralized mandates and gives a specific public example to anchor it.

Examples: Source 1

Dave Merkel states the tradeoff plainly: speed without accuracy is reckless, and accuracy without speed is useless.

• They argue that AI should earn autonomy through demonstrated outcomes rather than be handed the wheel on day one, and that humans should watch AI while AI watches humans—a requirement of the architecture, not a limitation.
• The post links to a published framework for deciding where AI actually belongs in a SOC and closes with MDR, cybersecurity, and security leadership tags.

A short argument built around a single tradeoff, followed by a link to a framework resource for practitioners.

For brands planning a campaign

A direct fit for a keynote, advisory engagement, or vendor webinar on AI in security operations, especially for an audience weighing how much autonomy to give AI in a SOC. The earned-autonomy and mutual-oversight framing gives a speaker a concrete, defensible structure for a talk.

Examples: Source 1

Ellingson cites Verizon breach data showing most breaches involve a human element and calls for structurally removing people from the attack surface. Merkel makes a parallel argument: because the adversary is human, defenders remain necessary, and AI changes the job rather than replacing it.

Ellingson builds long, source-cited extrapolations from threat reports toward a specific architectural thesis. Merkel's posts are shorter practitioner takes: a numbered prediction list and a policy reaction, with the supporting framework behind a link.

Read original · Original 2

Merkel argues that AI earns autonomy through demonstrated outcomes in a security operations center. Desai describes testing frontier models against real attack and defense scenarios, and the shift from scanning for known vulnerabilities to multi-step reasoning that builds exploit paths.

Desai writes from inside a vendor, framing the argument around a product launch and named partner programs. Merkel argues from the practitioner and policy side, without a product to position.

Read original · Original 2

Merkel says defenders will have to use AI or lose ground. Geffen says more hiring will not close the gap because attackers now compound across reasoning chains and shared memory.

Geffen frames the answer around autonomous pentesting agents and a company built to supply them, using a news article and a conference talk as hooks. Merkel sticks to predictions and operating principles, without pitching a product.

Read original · Original 2

The threat-report post reads four 2026 sources — CrowdStrike's Global Threat Report, WEF's outlook, CISA's national critical functions list, and Verizon's DBIR — into three curves: attack volume up 89% year over year, breakout time down to 29 minutes, and 82% of intrusions using no malware.

• It then extrapolates ten years out in phases, ending with the claim that security architecture must change because layered inspection is computationally expensive and defenders stay reactive.
• The second post turns the human-element statistic into a structural thesis: if most breaches involve a person, the answer is not better training but removing humans from the attack surface.
• For a keynote built around one big, arguable idea, that is a closer fit than a five-call predictions format.

Both posts carry a product argument.

• The human-element post builds toward EKI, presented as eliminating a breach class rather than improving authentication.
• That is the author advocating a specific architecture, so a client seeking neutral analysis would need to manage the commercial framing.
• The ten-year extrapolation is labeled as the author's own, not a finding of the cited reports, so it works as provocation, not as a citable forecast.
• Dave Merkel's posts do not cover SOC operating decisions, such as where AI earns autonomy or how a team stages deployment.
• For advisory guidance on those decisions, ask for specific examples.
• Their posts do not publish speaking terms, format, or audience information.

Examples: Source 1 · Source 2

The launch post spells out what a modern SOC platform now covers: unified exposure and threat management in one place; inline telemetry across network, identity, endpoint, cloud, AI, and deception; specialized agents built from a decade of SOC and threat-hunting work; and closed-loop remediation that turns detection into action.

• The second post reports testing Anthropic and OpenAI frontier models against real attack and defense scenarios, with three concrete observations: multi-step reasoning lets AI construct exploit paths instead of only scanning for known CVEs; architectural context improves exploit accuracy; and context should be limited to grounding the model.
• For practitioners deciding what to buy or build, that operational specificity is more directly useful than a predictions list.

Both posts are company communications.

• The first is a launch announcement for Zscaler's Agentic Security Operations solution, written as a thank-you to customers, teams, and partners; its substance is a product line rather than independent judgment.
• The second promotes a vendor blog.
• A client seeking a stated position on how much autonomy AI should get in a SOC would have to draw it out: the posts set no limits, oversight rules, or failure modes.
• They also offer no 2026 threat prediction or policy content, so the client would need to supply the keynote spine.
• Their posts do not publish speaking terms, format, or audience information.

Examples: Source 1 · Source 2

The Black Hat post describes OpenAI's agents sharing a memory system, building on each other's discoveries, chaining exploits, finding zero-days, and compromising real systems during an internal evaluation.

• It frames the real shift as agents gaining the ability to collaborate, reason, and execute complex attack paths together.
• The other post argues that hiring more CISOs will not fix the “bug-pocalypse”: AI writes more code faster with more vulnerabilities, the same AI is getting good at finding and exploiting them, and a twenty-person security team cannot manually review what AI coding assistants ship in a single afternoon across a thousand engineers.
• That staffing argument is a harder-edged answer to the same AI question Merkel addresses, and it will land differently with a technical audience than “the jobs change rather than disappear.”

Both posts point to Novee Security's AI pentesting agents as the answer, making this product-adjacent advocacy rather than independent analysis.

• The scope is narrower than a program covering SOC operating model, detection engineering, and the wider 2026 threat landscape: it focuses on offensive security and AI agents.
• The staffing argument may also be a poor fit for a room full of SOC managers, because its conclusion is that more headcount is insufficient.
• There are no predictions or policy segment, so a client seeking those topics would need to develop them elsewhere.

Examples: Source 1 · Source 2