Dan Arias is a software engineer focused on security, privacy, and digital identity. Their writing explores how to secure AI systems and why experienced engineering judgment matters when using AI to build software.
A named practitioner voice on securing AI agents and human oversight.
Best for
An AI-security platform wanting thought leadership on agent identity and human accountability.
Not for
A campaign needing guaranteed distribution into named buyer segments.
Key takeaways
AI agents should be identified, tracked, audited, and protected.
AI cannot replace the engineering judgment needed to build and maintain software safely.
A few models writing everyone's code create a concentration risk.
Check first: Confirm whether they accept sponsorships and have audience data on security buyers.
Main ideas
Engineering judgment is what AI cannot supply · Every AI agent should be identified, tracked, audited and protected · LLMs are probabilistic autocomplete; the labour model behind them is the problem
Dan Arias writes about software engineering, security, privacy, and digital identity, focusing on what responsible AI adoption demands from people and organizations.
They argue that AI tools can assist engineers but cannot replace the experience and judgment needed to build and maintain software safely.
Follow their work for a security-minded perspective on AI that connects technical choices to reliability, privacy, and the people affected by automation.
Arias makes broad AI debates tangible through engineering risks and human consequences.
In a post on AI-generated code and engineering judgment, they point to the many disciplines involved in shipping software—from architecture and accessibility to cybersecurity—and ask who will fix faulty code during a production incident.
In a post on securing AI agents, they call for agents to be identified, tracked, audited, and protected, extending familiar security principles to emerging systems.
They also question the costs of relying on a few shared AI models and absorbing workers’ knowledge while replacing workers (their post on AI and the workforce).
The focus is on who can judge AI’s output, secure it, and take responsibility for its effects.
Dan Arias writes about artificial intelligence as a working software engineer with a security and identity background.
Read more
Across the posts, the recurring claim is that AI is useful but cannot replace engineering judgment: models produce confident, plausible output, and experienced people with taste have to check it, reject what is wrong and keep flawed code out of production.
Alongside that is a security argument: AI agents should be identified, tracked, audited and protected like any other actor, with security built in from the start rather than bolted on.
Arias is also sceptical of replacement stories: they describe LLMs as probabilistic autocomplete, workers' knowledge as being absorbed rather than compensated, and a few models writing code across many companies as a shared dependency worth worrying about.
These positions fit together because they draw the line at supervision — models are good tools in the hands of good engineers, not autonomous ones.
Engineering judgment is what AI cannot supply
Dan Arias argues that claiming AI lets anyone build, debug and ship software is like claiming AI lets anyone practise medicine.
Read more
Software engineering, in this account, is broad and deep: distributed systems, architecture, continuous delivery, accessibility, performance at scale, data management and integrity, and cybersecurity all require experience, knowledge and taste.
LLMs can absorb information and return an answer wrapped in jargon convincing enough to trigger imposter syndrome, but they lack the experience to smell a bad answer.
Someone still has to demand evidence, verify it and stop bloated or underdeveloped code reaching production.
Imagine a hospital director announcing that anyone can now treat patients and perform surgery with ChatGPT: outrage, protests and lawsuits. Then consider the practical version of the same failure: an LLM writes or breaks something, non-engineers keep prompting it, and a software engineer is paged at 2 a.m. to fix it.
The post concedes the tool position rather than rejecting AI outright: LLMs are good tools under the supervision and handling of good software engineers. Dan Arias argues about who supervises, not whether to use the tools.
Every AI agent should be identified, tracked, audited and protected
Dan Arias treats security and identity as the lens for AI adoption.
Read more
Writing after leaving Auth0/Okta, they say their focus has stayed on security, privacy and identity, and that they spend free time on AI engineering with a strong emphasis on security and identity governance.
The stated goal is a world where every AI agent can be identified, tracked, audited and protected.
AI now sits inside software build and delivery pipelines and on both the defensive and offensive sides of cybersecurity, making trust and safety especially important.
Arias concludes that AI must be secure by design and that no target is too small: any weak point can be a stepping stone to a bigger one.
In the same post, Arias prefers passkeys and security keys as the best way to log in, a familiar case of identity done properly.
LLMs are probabilistic autocomplete; the labour model behind them is the problem
Dan Arias calls an LLM a probabilistic autocomplete with no genuine reasoning — efficient at copying humanity's homework, but unable to do its own.
Read more
The argument then turns to work: forcing employees to run their knowledge through these tools so the models absorb it is vampiristic.
Arias links the prospect of models trained on every message, document, commit and review from junior to principal engineers to mass unemployment.
Replacing tech recruiters with LLMs is a red flag, they say, because recruiting depends on authentic human-to-human engagement.
They also criticise marketing models as human-like intelligence for harming mental health and leaders' decisions.
The recruiter example and the post's closing statement: Dan Arias says they have loved AI and robots since childhood, but only when they help humanity rather than destroy it. They also like greenery, clean air and water, and animals, especially bees and birds.
The hostility is aimed at replacement and anthropomorphic marketing, not the technology. The same posts record hands-on work in AI engineering, keeping the critique consistent rather than contradictory.
A few models writing everyone's code create a concentration risk
A shorter argument about dependency: engineers required to use LLMs at work largely rely on the same handful of systems — GPT, Claude and Gemini — and inherit the same limitations, bugs, hallucinations and frustrations.
Read more
Their code ends up with the same defects.
Dan Arias finds it unsettling that a few systems write code across companies worldwide and asks what happens when one turns that accumulated know-how toward entering another company's business.
Organisations are sharing more than code with the models: knowledge about security, marketing, human resources and sales goes in too.
The post asks when one of these models might claim it can do what a business does and put a group of companies out of business. Dan Arias describes the situation as training the models daily to surpass everyone.
Learning in public and pointing to worthwhile resources
Alongside the criticism, Dan Arias makes continuous learning a priority and often shares what they find.
Read more
They spend free time levelling up on AI engineering, focusing on security and identity governance, to help people interact safely and securely with AI. They also recommend other people's technical material when it is substantive.
A short post recommends the IBM Technology YouTube channel, praising its cybersecurity series and, more recently, its AI series, with a link. Arias recommends the series as a good place to learn.
Answers a security buyer’s first question: what is this person working on now?
Read more
It offers the shortest route into the AI-security position that runs through the rest of their writing.
Arias says they have left Auth0/Okta but stayed in security, privacy and identity.
They now spend free time on AI engineering, focusing on security and identity governance.
The post repeats an Auth0 CISO’s line — "Security is part of the MVP" — and argues that it matters more now that AI sits inside build pipelines and both attack and defense.
They want a world where every AI agent can be identified, tracked, audited and protected, with security designed in from the start.
They close by saying they still favour passkeys and security keys for login.
For brands planning a campaign
A directly usable thesis for a security vendor: agent identity, tracking and audit as a product category. A partnership could have Arias argue that agent identity belongs in the same conversation as human identity, using passkeys as the familiar anchor.
Answers "where does this person stand on AI replacing engineers?"
Read more
It gives the fullest version of the argument, spelling out the specific disciplines and failure scenario.
Arias rejects the claim that AI lets anyone become a software engineer, comparing it to saying AI lets everyone practise medicine — imagine a hospital director announcing that anyone can now perform surgery with ChatGPT.
The post lists what the job actually involves: distributed systems, software architecture, continuous delivery, accessibility, performance at scale, data management and integrity, and cybersecurity.
In their account, large language models can produce plausible, jargon-heavy answers but lack the experience to spot a bad one.
A human has to prompt for evidence, check it and stop bloated code reaching production.
The post closes with the image of a 2 a.m. page to a real engineer when something breaks.
For brands planning a campaign
Strong raw material for a vendor whose message is augmentation, not replacement. The surgery analogy, list of engineering disciplines and 2 a.m. incident are concrete enough to reuse in a campaign aimed at engineering leaders who distrust the "anyone can build" pitch.
Answers "what risks does this person flag about company-wide AI adoption?"
Read more
The focus is concentration and dependency, not a tooling review.
A short post noting that engineers across companies now share the same handful of AI co-workers — GPT, Claude and Gemini — and therefore the same limitations, bugs and hallucinations.
Arias finds it unsettling that a few systems write code across many companies and asks when one could use that accumulated know-how to move into what another business does.
Companies share more than code: knowledge about security, marketing, human resources and sales is going into the same models.
For brands planning a campaign
Offers a security or platform vendor a dependency-risk angle: model concentration as a supply-chain concern. Treat it as a stated worry, not a demonstrated finding: the post raises the scenario as a question and presents no evidence for it.
Answers "what does this person actually want from AI?"
Read more
This is the values piece and sets the boundary on which arguments will land with them.
Arias calls an LLM a probabilistic autocomplete with no genuine reasoning, good at copying humanity's homework but not doing its own.
They describe forcing workers to feed their knowledge into models as "vampiristic," and point to tech recruiters being replaced as a warning sign, because recruiting needs genuine human engagement.
They also say the marketing of LLMs as human-like intelligence is damaging to mental health and to leaders' decisions, and close by saying they have liked AI and robots since childhood — but for helping humanity, not replacing it — and that they like greenery, clean air, water, and animals, especially bees and birds.
For brands planning a campaign
Read this before pitching.
Arias is sceptical of replacement stories and anthropomorphic model marketing.
A brief framed around AI that supports and extends skilled workers fits; a "do more with fewer people" pitch is likely to be rejected.
Answers "what does this person put in front of their audience?"
Read more
It is the thinnest item here, included for completeness rather than depth.
A two-sentence post recommending the IBM Technology YouTube channel and praising its cybersecurity series and, more recently, its AI series, with a link attached. It offers no commentary on specific episodes or claims.
For brands planning a campaign
Shows Arias will amplify substantive, vendor-produced technical education. A company that already publishes a strong explainer series has a plausible, low-friction reason to approach them; a share does not imply a partnership or payment.
Continuous learning and curated technical resources2
AI is not a shortcut to software engineering1
Security and identity for AI agents1
AI is not a shortcut to software engineering
Dan Arias argues that LLMs cannot make anyone a software engineer, comparing that claim to saying AI lets anyone practice medicine.
Read more
They point to distributed systems, software architecture, continuous delivery, accessibility, performance at scale, data management and integrity, and cybersecurity as fields that require experience, knowledge, and taste. In their view, LLM output can sound authoritative and jargon-heavy while hiding errors that only experienced engineers can spot.
An extended analogy to medicine leads into a practical list of engineering disciplines.
The argument presents LLMs as useful tools under the supervision of good software engineers, not as something to reject outright.
For brands planning a campaign
A security company can use this well-argued case for human oversight in AI security content, advisory work, or thought leadership on reviewing model output, incident response, and production safeguards.
Reflecting on moving on from Auth0/Okta, Dan Arias says their focus remains on security, privacy, and identity, and that they are spending free time on AI engineering, with a strong focus on security and identity governance.
Read more
They want every AI agent identified, tracked, audited, and protected. They reiterate that security is part of the MVP and call for AI to be secure by design.
A first-person update sets out short principles and a clear desired outcome for AI agents.
They also state an ongoing personal preference for passkeys and security keys.
For brands planning a campaign
Strong fit for a security vendor's campaign on agent identity, auditability, and AI security posture. A collaboration could center on secure agent deployment, identity governance, or passkeys, building on the creator's existing position.
Across two posts, Dan Arias argues that LLMs are probabilistic autocomplete without genuine reasoning, that companies are training models to surpass the workers whose knowledge they absorb, and that replacing people, including recruiters, with LLMs is destructive.
Read more
They warn that a few models writing code across many companies could turn shared business knowledge into a competitive risk, and point to mental-health and environmental costs.
Rhetorical questions, personal ethical framing, and a dystopian scenario make the argument concrete.
They say they love AI and robots and want technology that helps humanity, not destroys it.
For brands planning a campaign
This gives a security brand a clear, human-centric angle for content about responsible AI, workforce impact, and trust. Automation framed as a simple replacement for human judgment would conflict with the creator's stated position.
Continuous learning and curated technical resources
Dan Arias points readers to IBM Technology's YouTube channel for cybersecurity and AI series and describes spending free time leveling up on AI engineering, with a focus on security and identity governance.
Read more
Their posts show a pattern of learning in public and recommending useful resources.
A resource recommendation paired with a note about learning in public.
For brands planning a campaign
A fit for educational partnerships, resource roundups, and explainer content. Dan Arias brings a peer learner's and curator's perspective, not just a critic's, making them a natural partner for technical education content.
The posts are first-person LinkedIn essays built around a single claim, an analogy or rhetorical question, and practical examples.
Read more
Dan Arias writes in a direct, opinionated voice and often closes with a values statement rather than a neutral summary.
A first-person essay using analogy, rhetorical questions, and values-based framing.
For brands planning a campaign
A collaboration should leave room for a point of view and reasoning rather than a scripted product mention. Their short, text-based arguments and explainers can work as thought-leadership or educational posts for a security brand.
Dan Arias's writing documents one workplace tie: a past role at Auth0/Okta, referenced when they say they are "no longer part of the Auth0/Okta ship" while staying engaged with security, privacy, and identity.
Read more
The only other named business is IBM, whose Technology YouTube channel they recommend. Ask directly whether any commercial relationship exists.
Auth0 / Okta
Employment
Arias writes that they are "no longer part of the Auth0/Okta ship" but remain involved with security, privacy, and identity work.
Read more
The post recounts advice from a CISO during that period and describes the role as a past one, not a current engagement.
Original disclosure
While no longer part of the Auth0/Okta ship, my heart and mind continue to be engaged as much as possible with security, privacy, and identity.
Arias calls the IBM Technology YouTube channel "a gold mine of technical resources" and says they follow its cybersecurity series and have moved into its AI series.
Dan Arias writes short, argument-first posts about artificial intelligence from the perspective of a working software engineer with a security and identity background.
Read more
Read the position pieces in order: first, the argument that AI does not make anyone a software engineer because judgment, taste and verification still decide what ships; then, the case that every AI agent should be identified, tracked, audited and protected, with security built in from the start; and finally, critiques that call LLMs probabilistic autocomplete, question the labour model behind training them on employees' knowledge, and warn that a handful of models writing code across many companies creates a shared dependency.
A short recommendation of an external technical channel shows the same habit of learning in public.
Readers looking for technical education with a point of view, rather than neutral explainers, will find a clear through-line.
Both write about how AI is changing product and engineering work and what still depends on human judgment.
Read more
Rupinder Singh's posts report conversations with named practitioners about why AI product work is hard: the product is probabilistic, yet the manager remains accountable. Dan Arias raises the same concern about confident model output that someone has to review.
Rupinder Singh's material centres on interviews and podcast episodes with others about the product manager's role. Dan Arias writes first-person essays on engineering practice and security, arguing a critique rather than hosting a conversation.
Both examine how agentic coding is reshaping software teams and which skills will last.
Read more
Andrew Ng's posts describe AI-native teams, engineers taking on product decisions, and deciding what to build becoming the bottleneck once building gets cheap. That overlaps directly with Dan Arias's focus on where engineering judgment matters most.
Andrew Ng frames the change as expansion and opportunity, backed by conferences, curriculum questions and hiring trends. Dan Arias is the sceptic in the pair: verification, incident risk, agent identity and the limits of model reasoning matter more than the upside.
Both test what LLMs can and cannot do in day-to-day systems, and write direct first-person analysis for practitioners rather than general-audience commentary.
Read more
Noah Stambovsky's point that LLMs can join data across disconnected tools but cannot reconcile data that disagrees with itself echoes the practical limits Dan Arias identifies in engineering model output.
Noah Stambovsky writes about go-to-market operations, CRM and data consolidation, including HubSpot survey findings and operational handoffs. Dan Arias focuses on software delivery and security, with identity governance and production incidents as recurring examples.
No partnership or sponsorship offer appears in Dan Arias's posts.
Read more
The five posts offer personal commentary on software engineering, AI, security, and identity; none announces a sponsorship policy, media kit, rate card, or business inquiry route.
They take an engineering and security-practitioner perspective rather than presenting creator placements, so a brand would be starting a conversation rather than choosing from published packages.
Ask directly whether they accept sponsorships and, if so, what formats and terms they offer; confirm those details before building a campaign plan around them.
Brief A: an AI-security platform wants a named practitioner voice to lead a short thought-leadership series on securing AI agents and keeping humans accountable for model output — opinion essays, one live session, and advisory input on how the vendor talks about agent identity.
Read more
This campaign fits Dan Arias's posts closely, for specific reasons.
Their clearest current position is that AI agents need identity treatment: they want "a world where every AI Agent can be identified, tracked, audited, and protected," restate the line "Security is part of the MVP," and argue that security has to be designed in from the start rather than bolted on once AI sits inside build pipelines and both attack and defense.
A security vendor can adopt that argument without inventing a stance for the creator.
It is also anchored, not abstract: the same post closes on a concrete login preference — passkeys and security keys — giving the campaign a familiar, human-identity hook for agent identity.
Their posts also argue the human-oversight half of the case.
The software-engineering post rejects the idea that AI makes anyone an engineer, lists what the job actually involves (distributed systems, architecture, continuous delivery, accessibility, performance at scale, data management and integrity, cybersecurity), warns that model output can be jargon-heavy and wrong, and lands on the 2 a.m. page to a real engineer when generated code breaks in production.
For a vendor whose message is augmentation rather than replacement, that is a ready-made supporting argument with a worked example.
Practical fit: the observed format is a short first-person argument — a single claim, an analogy or rhetorical question, concrete examples, and a values statement at the end rather than a neutral summary.
It suits a campaign that wants a point of view, not a scripted product mention.
They have also directed their audience to vendor-produced technical education, recommending IBM Technology's cybersecurity and AI series as a resource.
That was a share, not a paid partnership, so it signals appetite for substantive technical content, not proven commercial interest.
The audience section records roughly 1,500 LinkedIn followers but no impressions, engagement, geography, seniority, or job-function data.
Ask for post-level analytics and audience demographics before relying on reach into named buyer segments.
The posts also do not specify sponsorship terms or an inquiry route, so negotiate scope and fees from scratch.
When to choose a specialist
Brief B: a security vendor wants a paid awareness campaign with guaranteed distribution into named buyer segments — an agreed number of enterprise CISOs or security decision-makers in specified regions — planned against published sponsorship packages and a rate card.
Read more
This campaign is not a fit based on Dan Arias's posts, for two separate reasons.
First, the audience cannot be verified as specified.
The only measured figure is about 1,500 LinkedIn followers, with no impressions, engagement counts, follower growth, geography, seniority, or job-function data.
Posts that appeal to security-minded engineers do not show that such engineers are the audience.
A planner who needs a segment guarantee has no information here to size it, price it, or promise a result against it.
Before signing, ask for recent post-level analytics and platform demographic data or an audience survey.
Second, the commercial format is undocumented.
The posts contain no sponsorship announcement, rate card, media kit, or business inquiry route.
A campaign plan cannot rely on packages that have not been shown to exist, and treating a personal resource recommendation as evidence of a paid placement would misread the posts.
There is also a message mismatch worth flagging separately, because it can sink a campaign even if the commercial terms are agreed.
A narrative built on doing more with fewer engineers, or on AI replacing human judgment, runs directly against the writing: the argument that LLMs are "probabilistic autocomplete" with no genuine reasoning, the description of feeding workers' knowledge into models as "vampiristic," the warning that recruiters being replaced is a red flag, and the statement that technology should help humanity rather than destroy it.
A campaign whose central claim is headcount reduction would ask them to argue against their own published position.
The audience section also offers no evidence that the following matches a security vendor's target buyers.
A worked campaign example
Worked angle for Brief A: "Agents need identity, not just access."
The campaign makes the vendor's argument through a claim Dan Arias already publishes: AI agents should be identified, tracked, audited, and protected the same way people are, and security belongs in the build from the start rather than in review afterwards. The vendor need only supply the technical detail, review, and distribution — not the thesis.
Three beats, each carrying one reason, one example, and one action:
1. Why agent identity is identity work. Open on the line that security is part of the MVP, applied to a world where AI sits inside build, delivery, and both sides of the defense equation. Concrete example: an agent acting on a system with no durable identity trail presents the same class of problem as an unmanaged service account. The vendor's engineers can explain the failure mode in specific, checkable terms. 2. Why humans still own the output. Use the model-output argument already in the posts: plausible, jargon-heavy answers that sound authoritative and can be wrong, and the operational reality that a real engineer gets paged when generated code breaks. Example: the prompt-for-evidence-and-verify loop as an actual review step, not a slogan. Action for the reader: name who is accountable when an agent ships something bad. 3. What secure-by-design looks like in practice. Connect the agent-identity argument to the vendor's product category, using passkeys and security keys as the familiar human-identity anchor the creator already favours. Action: a short checklist an engineering leader can run before deploying an agent with production access.
Format: first-person LinkedIn essays in the creator's observed register — one claim each, an analogy or pointed rhetorical question, concrete examples, and a values statement at the close rather than a neutral summary. Keep product claims to what the vendor's own engineers will stand behind; the creator's credibility rests on being read as a practitioner, not a spokesperson.
Order of work: agree scope, fee, and disclosure first, since no published package exists; run the three essays or an equivalent series; then decide on a live session only if the written work lands.
Before contracting, confirm whether the vendor may republish or excerpt the material and who signs off on product assertions. Agree to share post-level analytics from the first deliverable so the second campaign decision rests on data rather than follower count.
Consider Andrew Ng when the vendor wants the widest association with the "future of software engineering" story rather than a security specialist.
Read more
The posts argue that AI is accelerating software engineering more than any other profession, that deciding what to build is now the bottleneck rather than the building, that engineers are absorbing product work and product managers are learning to build, and that design, marketing, and legal compliance become the slow steps once coding speeds up.
An AI Dev conference on April 28-29 in San Francisco is named as the venue for that discussion.
For a vendor whose agent-identity message needs to sit inside a broader "how teams and work change with AI" narrative, this brings an established event platform and a team-organization argument the campaign can ride on.
The posts cover engineering productivity, team shape, and hiring, not security or identity.
Dan Arias makes no agent-identity, audit, or human-accountability argument for the vendor to borrow, so the vendor would have to supply the security thesis and would pay mainly for presence.
The register is measured and explanatory, not the direct first-person argument the campaign needs.
One of the two posts is cut short; ask for the complete post before evaluating the comparison.
Consider Rupinder Singh when the vendor wants to reach product managers who choose and ship AI features, using an interview or podcast format instead of written essays.
Read more
The posts are built around practitioner conversations and frame AI product management as a decision problem under uncertainty — the cited point is that the hardest part is not picking a model or writing a prompt but making product decisions when the product is probabilistic and the manager is still accountable. For a vendor seeking a co-produced episode aimed at PMs evaluating AI tooling, that format matches the observed output more closely than opinion essays.
The subject is product management, not security or identity, and both posts promote conversations with other people rather than stating Dan Arias's own security position.
The vendor would depend on those guests to carry the security argument, losing the practitioner-voice quality the campaign needs.
Ask for audience size, composition, and commercial availability before comparing reach.
Consider Noah Stambovsky when the agent message is better framed through integration and data reality than through identity and accountability — getting agents to work across many disconnected tools without consolidating first.
Read more
The posts argue that with MCPs, LLMs can produce join-like reporting across a dozen separate tools, that the unresolved limit is data that actively disagrees with itself, and that teams which stayed siloed with clean data may run more smoothly than those that consolidated badly.
A second post uses a 1,200-respondent HubSpot survey to press on lifecycle definitions, handoffs, and ownership inside a CRM.
For a vendor selling agent tooling into messy go-to-market stacks, that is a more useful argument than agent identity on its own.
This shifts the campaign's centre of gravity from security to marketing and sales operations and CRM plumbing.
Dan Arias's posts say nothing about agent identity, audit trails, or who is accountable when a model ships something wrong, so the vendor would have to supply the security thesis.
This is the better choice only if the buyer decision is really about GTM data integration, making it a different campaign from one aimed at security and identity.
This profile is built on five LinkedIn posts attributed to Dan Arias's account at linkedin.com/in/getdanarias, each held as full post text and reviewed in full: four argument essays on AI, software engineering, security and identity, and one short recommendation of an external technical channel.
The posts carry no visible publication dates, so the ideas are grouped by subject rather than presented as a timeline, and no change over time can be traced from this material.
Account details, including the roughly 1,500 LinkedIn followers cited in the audience discussion, were observed on October 9, 2026.
Comparator material comes from two posts each by Rupinder Singh, Andrew Ng and Noah Stambovsky, dated between April 13 and September 30, 2026.
The set contains no long-form articles, talk recordings, video transcripts, code, audience demographics, engagement data, sponsorship packages or rate cards, so positions and commercial formats beyond these short-form posts remain open gaps that would need direct evidence before a campaign plan depends on them.