Tori Adams

The remedy is consistent across posts: a distinct identity for every production agent, a named owner, permissions limited to a specific task, an independent log of what the agent did, and a clear way to revoke access. The same controls apply to agents that run unattended or pass work to one another.

In “Your AI agent should not inherit your digital badge,” Adams notes that organizations give agents access to email, financial records, source code, and cloud infrastructure through credentials borrowed from people, so the log may say the employee acted.

• The badge post spells out the requirements: identity, owner, task-specific permissions, audit trail, revocation path.
• A later post on always-on agents adds the operational questions: what the agent can reach, what it may change, and who can stop it.
• A post on agents communicating in ways people cannot read recommends enforcing permissions outside the model and logging actions independently.

Read original · Original 2 · Original 3 · Original 4 · Original 5

Restricting who may supply a system is a procurement and national-security question; establishing what a system has demonstrated is a testing question.

• Acquisition documents, Adams argues, should set the evidence standard.
• Today, the default evidence for autonomous performance is often the vendor’s own claim.

In a counter-UAS post, Adams traces Section 1709 of the FY2025 NDAA and the FCC’s December 23, 2025 Covered List action, including DJI and Autel petitions and Ninth Circuit litigation.

• Adams contrasts that settled supply-chain regime with vendor-reported hit probabilities for kinetic drone interceptors, ranging from 50 to 100 percent.
• Adams calls a fifty-point spread the absence of a test record and points to DoD Directive 3000.09, which requires autonomous and semi-autonomous weapons to be tested against realistic conditions and adaptive adversaries, as well as FY2026 NDAA provisions on the authority side.

Read original · Original 2 · Original 3

The immediate danger, they argue, is practical: industrialized attacks against defenses that remain artisanal, agents with excessive permissions, and, at larger scale, swarms whose behavior emerges from agents interacting rather than from any single design.

• Adams describes opacity in agent communication as an oversight problem, not proof of bad intent; in one post, they say plainly that opacity alone does not establish malicious intent.
• Their conclusion is that safety and deployment speed are not opposing goals: bounded authority and testing make autonomy usable.

Adams cites a CrowdStrike report describing an attacker who used an open-source agentic penetration-testing tool alongside large language models against South Korean financial organizations.

• Configuration files and coding-session histories were found on attacker-controlled infrastructure.
• Adams draws four pre-deployment requirements: distinct agent identity, recorded tool calls and actions, human approval for consequential actions, and tested containment.
• They later extend the question to Anthropic’s work on systems that would let agents operate microscopes, lasers, liquid handlers, and robotic arms: what is an agent allowed to do, and what happens when it is wrong?

The focus broadens over the reviewed period.

• Early posts address digital actions such as email, databases, and business processes.
• A post on Anthropic’s equipment-control work marks a shift to physical actions, where permissions, fail-safes, and human oversight become operational concerns.

Read original · Original 2 · Original 3 · Original 4

They draw three conclusions.

• The infrastructure build-out in data centers, chips, power, and cooling may be inflationary before it becomes deflationary.
• Capital markets are pricing a productivity revolution that has yet to appear in economy-wide productivity, putting companies in a “prove-it phase” in which they must show measurable productivity, revenue, risk reduction, or mission improvement.
• Finally, the model is not the constraint: fragmented data, legacy systems, weak governance, unclear authority for agents, and procurement processes designed around human labor will decide the results.
• Applying powerful models to those conditions can produce expensive dysfunction rather than productivity.

Adams describes an ultimate deflationary technology that creates significant inflation while it is built, and reframes the executive question from “Do we have an AI strategy?” to “Where can I show measurable productivity, revenue, risk reduction or mission improvement from AI?”

Read original

Written in the vocabulary of procurement and pre-deployment review, the lists are transferable. Adams also directs readers to their Substack for the longer version; ask for those essays if they’re relevant to your campaign.

One post draws four signals from reports on the Army and NSA: dedicated funding as a mission requirement, wargames linked directly to acquisition, agents trained and qualified for defined cyber work roles, and strategic competition as the driver. Adams then lists the questions a government organization must answer: who funds the agents, who trains and qualifies them, what authorities they hold, what needs human approval, how decisions are monitored and reversed, and who is accountable when an agent is wrong.

Read original · Original 2 · Original 3 · Original 4

The industries vary—automotive sourcing and battery regulation, food safety after large recalls, aerospace supply chains, semiconductor compliance under the CHIPS Act, and EU due-diligence and tariff risk—but the structure is consistent.

• One promotional post names customers; ask for substantiation if those references matter to your campaign.
• These posts differ from the federal-AI governance writing.

A manufacturing whitepaper announcement lists benefits such as automated compliance, circular product design, and new revenue from resale and repair. A food-safety post cites a USDA recall of 11.7 million pounds of ready-to-eat meat and a Listeria outbreak to argue for end-to-end traceability and faster recalls.

The emphasis shifts over the reviewed period. The earlier product-passport posts are promotional; the later posts are largely federal AI safety, governance, and acquisition analysis, written in Adams’s own voice and grounded in named statutes, directives, and news reports.

Read original · Original 2 · Original 3 · Original 4 · Original 5 · Original 6

Adams makes agent identity a concrete accountability problem: if an agent uses a person's credentials, the audit log can make its actions look like that person's. The post names practical controls: an agent-specific identity, owner, limited permissions, audit trail, and way to revoke access.

For brands planning a campaign

For federal buyers, this is a useful starting point for discussing agent authority and procurement requirements. A co-authored analysis could turn the listed controls into testable acquisition criteria.

Drawing on a report about Claude agents controlling laboratory and engineering equipment, Adams distinguishes systems that recommend actions from systems that take them. The move into physical equipment makes permissions, monitoring, fail-safes, and human oversight immediate operational questions.

For brands planning a campaign

This could anchor a campaign about assurance beyond software agents. A webinar could ask what evidence buyers need before an agent operates equipment and how they can verify that safeguards work.

Start with their high-level position, then read the more specific identity and assurance examples for implementation detail.

Adams rejects a choice between avoiding AI and adopting it without limits. They connect the need for AI in cyber defense with risks from agents and agent swarms, then name identity, bounded authority, monitoring, adversarial testing, and containment as practical controls.

For brands planning a campaign

This is a strong framing for a federal acquisition webinar: agencies may need AI to defend at machine speed, but deployments should demonstrate that operators can monitor and contain it.

Adams separates supplier restrictions from proof that a system performs safely. They contrast vendor-reported accuracy claims for kinetic drone interceptors with the need for testing in realistic conditions, and point to acquisition as a place where standards for proof matter.

For brands planning a campaign

For a campaign on autonomous-systems assurance, this offers a concrete acquisition case: distinguish who is permitted to supply a system from what testing demonstrates about its performance. Confirm the post's legal and factual details before using them in a public-facing campaign.

Borrowed logins create an accountability gap: when an agent acts, the audit trail names the person.

• Their fix is consistent across posts: a distinct identity for every production agent, a named owner, task-specific permissions, an independent action log, and a revocation path.
• The same frame appears in posts about agents that keep working unattended, agents that hand work to each other, and attackers who use agentic tools against financial institutions.

They turn the argument into short, numbered requirement lists for acquisition documents and pre-deployment checks, a structure that transfers directly to a joint article or webinar outline.

For brands planning a campaign

For a defense or identity-security vendor, this is the strongest co-authoring topic in the corpus: Tori Adams’s most repeated and developed argument, written in the language of procurement and acceptance testing used by government buyers.

Examples: Source 1 · Source 2 · Source 3 · Source 4

In a counter-UAS post, they walk through Section 1709 of the FY2025 NDAA, the FCC’s December 2025 Covered List action, and the DJI and Autel petitions and Ninth Circuit litigation.

• They contrast that supply-chain regime with vendor self-reported accuracy of 50 to 100 percent for drone interceptors—a spread they call the absence of a test record.
• They also point to DoD Directive 3000.09 on testing autonomous and semi-autonomous weapons and to FY2026 NDAA sections 912 and 1261 on the authority side.

For brands planning a campaign

For anyone selling assurance, test-and-evaluation, or counter-UAS capability to government, Adams offers a ready topic and close readings of statutes, directives, and petitions.

Examples: Source 1 · Source 2 · Source 3

Their CrowdStrike post describes an attacker using an open-source agentic penetration-testing tool alongside large language models, then turns the example into four pre-deployment requirements. A later post extends the question to Anthropic systems that would operate microscopes, lasers, liquid handlers, and robotic arms; another asks what happens to oversight when agents communicate in ways humans cannot read.

They are careful about intent: in a post on unreadable agent communication, Adams states plainly that opacity alone does not establish malicious intent.

For brands planning a campaign

A cybersecurity vendor can sponsor analysis that moves agentic AI out of futurism and into current-threat territory. Adams’s stated position—that safety and deployment speed are not opposing goals—gives a sponsor room to sell controls rather than delay.

Examples: Source 1 · Source 2 · Source 3 · Source 4

They conclude that the next bottleneck is the organization, not the model: fragmented data, legacy systems, weak governance, unclear authority for agents, and procurement barriers designed around human labor.

For brands planning a campaign

Enterprise transformation and management-consulting brands get a lane distinct from Adams’s federal-security writing. Their question—where can measurable productivity, revenue, risk reduction, or mission improvement be shown?—is a practical hook for a co-authored ROI piece.

Examples: Source 1

These company announcements use hashtag stacks, calls to read a Substack or blog, and product claims rather than the independent analysis found in the federal-AI posts.

Boeing, the US Department of Defense, and major US battery manufacturers appear as named customers in one promotional post. Ask for references before relying on those customer claims.

For brands planning a campaign

These posts offer industry-by-industry framing for supply-chain, compliance, and digital-product-passport vendors. Brands seeking editorial credibility should distinguish them from Adams’s AI-governance work.

Examples: Source 1 · Source 2 · Source 3 · Source 4 · Source 5 · Source 6

The Army and NSA post assembles four reports into four signals, then asks who funds, trains, qualifies, monitors, audits, and answers for the agents.

For brands planning a campaign

Pitch primary sources, not slogans.

• Adams’s posts stand or fall on named documents and concrete requirements.
• They consistently point readers to the Substack for the longer version, a natural place for a sponsored deep dive.

Examples: Source 1 · Source 2 · Source 3 · Source 4

Written in SIMBA's corporate voice — "At SIMBA, we're at the forefront of this evolution" and "our latest whitepaper" — they link to the company's blog and Substack. They do not state Adams's title or whether they are an employee, contractor, or partner; ask what their relationship with SIMBA Chain is.

Original disclosure

At SIMBA, we're at the forefront of this evolution, helping businesses unlock new levels of efficiency, transparency, and innovation.

View example

The newsletter develops arguments their shorter LinkedIn posts summarize.

Original disclosure

Check out my substack for more on this

View example

Their framing of agents as digital workers and argument that capability should not outrun controllability overlap directly with Adams's posts on badges and authority.

Sharma argues from governance principles, asking leaders open questions about incentives, regulation, and who benefits from tighter rules. Adams anchors each post in a named document — an NDAA section, an FCC listing action, a DoD directive, a vendor report — and focuses on acquisition and pre-deployment requirements rather than board-level framing.

Read original · Original 2

Their mining posts argue that remote and autonomous operation is proven rather than speculative, backing the case with operating history, returns, and deployment numbers. Both writers push past the demo and ask what an autonomous system has actually demonstrated.

Goldsmith writes as an operator building a commercial service, with paid-on-production incentives and an equipment-to-command-centre story. Adams writes about digital agents from a policy and procurement angle; their posts describe no comparable operating or commercial model.

Read original · Original 2

A conference recap quotes research showing executives measuring AI by individual productivity while few workers have real proficiency, and argues that ethics and effectiveness are not separate conversations — a close cousin of Adams's point that safety and deployment speed are not a trade-off.

Demirci's posts are event recaps and awareness campaigns framed around ethics frameworks, ISO 42001, and the EU AI Act. Adams writes short operational analyses built on statutes, directives, threat reporting, and acceptance testing, targeting federal agencies and defense buyers rather than a general governance audience.

Read original · Original 2

Sharma writes about the same practical problem Tori Adams addresses: what an AI agent should be allowed to access, decide, and do.

• They name the same controls—a clear purpose, a named owner, its own identity, an access boundary, a risk level, an approval process, and an audit trail—and argue that agents should be governed more like digital workers than treated as a productivity feature.
• A sponsor whose product story fits operating discipline for boards and CISOs better than acquisition criteria for contracting officers will find a closer fit here.
• Sharma also frames the frontier-AI risk debate as capability outrunning controllability and asks what powers should be delegated before control is proven, a useful hook for a co-authored piece aimed at enterprise decision-makers.

The Sharma posts lack the federal detail a campaign selling assurance to government would need.

• They cite no statutes, DoD directives, acquisition authorities, test-and-evaluation standards, or counter-UAS supply-chain material; a sponsor would need to provide that substance.
• The two posts, published in August 2026 and September 2026, are opinion essays that point readers to a longer article rather than presenting it.
• Ask for the full article before judging its depth.

Examples: Source 1 · Source 2

Goldsmith's posts argue that autonomous and semi-autonomous mining is already proven, describe one operator running equipment across multiple mines from a surface command center, cite internal rates of return for conventional, local-surface, and centralized-command operations, and pitch a commercial model in which the supplier is paid when material moves rather than when equipment arrives.

• For a sponsor whose autonomous-systems assurance pitch centers on physical equipment, production risk, and outcome-based contracts, that operating experience speaks directly to the buyer's concerns.
• Tori Adams's posts, by contrast, stay in the digital and cyber domain.

This campaign centers on a single company's platform and commercial model, not federal AI governance.

• The two promotional posts, published in June 2026, contain no agent identity, procurement criteria, cybersecurity, federal acquisition, or test-and-evaluation content.
• A sponsor seeking the government acquisition audience Tori Adams addresses would need another channel.

Examples: Source 1 · Source 2

Demirci writes about AI governance through named frameworks, including ISO 42001 and the EU AI Act, and presents governance as a way to protect trust, privacy, and security rather than as paperwork.

• One post recaps a London AI governance summit and examines ethics frameworks that build participation into the AI lifecycle, alongside a data point that executives measure AI success by individual productivity while only a small share of knowledge workers have real AI proficiency.
• The other argues that shared images can be reused through deepfake technology and that both awareness and regulation matter.
• A sponsor selling compliance or governance tooling to EU or multinational organisations will find closer alignment with this vocabulary and audience than in Tori Adams's federal-first writing.

The posts lack the federal acquisition detail Tori Adams covers: United States procurement, statutory citations, acquisition authorities, and concrete test-and-evaluation standards.

• Agent authority and assurance appear as principles, not criteria a buyer can score.
• The two posts—a summit recap and an awareness piece, published in August 2025 and November 2025—are older and fewer than Tori Adams's run of agent-governance and counter-UAS writing.
• Ask for scored evaluation criteria before treating them as a fit for an acquisition-focused campaign.

Examples: Source 1 · Source 2