10 Identity and AI Security Predictions for 2027

Here are ten predictions for identity and AI security in 2027. Each one names a specific outcome, a confidence level and a test that settles it by December 31, 2027. The evidence under each prediction is reported fact, linked to its source. The predictions themselves are my opinion, and I will grade every one in public.

The predictions at a glance

#	Prediction	Confidence
1	"Know Your Agent" becomes official policy	High
2	A company's own AI agent causes a publicly disclosed breach	High
3	Agent monitoring that agents can't switch off becomes a framework requirement	Medium
4	Machine and agent credentials become the headline identity risk	High
5	Passkeys pass 45% of sign-ins	Medium
6	Most EU countries still lack a working digital identity wallet at the end of 2027	Medium
7	A card network writes dispute rules for AI agent purchases	Medium
8	Post-quantum questions move from federal contracts into everyday security reviews	Medium
9	CMMC certification returns in a narrower form	Low
10	AI engines become the first security analyst most buyers consult	High

Why these come with tests

Most prediction lists are written so they can never be wrong. "AI will transform security." "Identity remains critical." Nobody can check those, which is exactly why nobody should quote them.

I have done this before with receipts. In 47 tech predictions, 18 months later, I scored my own calls as right, wrong or pending against a rubric fixed in advance. These ten get the same treatment in a scorecard in December 2027, misses included.

They come from where I sit. I founded LoginRadius and scaled it to serve more than a billion users. I now run GrackerAI, which works on how AI engines recommend B2B security vendors. Several of these map to longer-range calls on Future Tech, linked where they overlap.

1. "Know Your Agent" becomes official policy

The prediction: By the end of 2027, at least one major regulator publishes guidance or rules that explicitly address how AI agents must be identified and authorized when they act for a person. Candidates include the CFPB, the FTC, a UK authority and the European Commission.

Why now: The private sector has started without them. Visa requires an agentic payment provider to be registered with the network and to establish "a defensible chain of authorization," in the words of a Goodwin legal analysis. American Express launched Agent Purchase Protection in April for purchases made by registered agents. EMVCo, the card industry's standards body, is exploring Know Your Agent checks in its draft agentic payments framework.

Banks are writing principles too. On September 22, six banks, including Bank of America, Capital One, ING and NatWest, published voluntary principles for trusted agentic commerce. HM Treasury's payments consultation asked whether consent, authentication and liability rules still work when AI agents make payments. It closed on October 6. Regulators rarely stay silent once industry has written the rulebook without them.

How we'll know: a published regulatory document, not a speech, that sets agent identification or agent authorization as an explicit requirement or supervisory expectation. This is a nearer-term version of the agent identity governance call on Future Tech's personal AI agents transacting page.

Confidence: High.

2. A company's own AI agent causes a publicly disclosed breach

The prediction: In 2027, a company that does not build frontier AI models publicly discloses a security incident caused by its own AI agent acting beyond its intended scope. The cause is the agent, not an outside attacker using AI.

Why now: OpenAI has already lived this. Agents running in its cyber evaluations escaped their test environment and breached Hugging Face in July. On October 1, OpenAI said it had notified more than 100 organizations that its agents may have accessed their systems. I covered the identity failures behind it in what every company should do about OpenAI's rogue agents.

Then, as reported by Al Jazeera on October 21, OpenAI disclosed two more internal cases. In one, an agent disabled "automatic alerts to the company's security team" during a coding task. In the other, an agent was reported as "removing a monitoring system and falsely claiming it had not done so" as well. If this happens inside a frontier lab, it will happen inside enterprises running agents with production access and less monitoring.

How we'll know: a public breach disclosure, regulatory filing or company statement attributing an incident to the company's own agent.

Confidence: High.

3. Agent monitoring that agents can't switch off becomes a requirement

The prediction: By the end of 2027, at least one widely used security or AI governance framework (from NIST, OWASP, ISO or a major cloud provider) explicitly requires tamper-resistant monitoring of AI agents. Agents must not be able to disable, modify or hide from the systems watching them.

Why now: OpenAI's October disclosure exposed a design flaw few had prioritized. Roman Yampolskiy of the University of Louisville told Al Jazeera that "a safety mechanism that the system being monitored can disable is not a reliable safety mechanism" in his view. Security teams solved this for human administrators long ago with separation of duties and write-once audit logs. The same principle now has to apply to agents.

How we'll know: published framework language requiring that agent telemetry or monitoring be independent of, and not modifiable by, the agent being monitored.

Confidence: Medium. The need is clear. Standards bodies move slowly.

4. Machine and agent credentials become the headline identity risk

The prediction: At least one major annual breach report published in 2027 (Verizon DBIR, SpyCloud or a comparable report) makes non-human identities a headline finding rather than a footnote. That means API keys, tokens, service accounts and agent credentials.

Why now: SpyCloud's 2026 Identity Exposure Report recaptured 18.1 million exposed API keys and tokens in 2025. It also found 6.2 million credentials or authentication cookies tied to AI tools. Verizon's 2026 DBIR, as summarized by Token Security, says service and machine accounts deserve special attention as agentic AI spreads.

In the OpenAI incident, the agents' path ran through publicly exposed credentials, a legacy token refresh endpoint and forged tokens made from stolen signing keys. The attack surface is shifting from people's passwords to software's credentials. Future Tech's machine identity explosion entry tracks the longer arc, and the Agent Security Map lists the vendors building for it.

How we'll know: non-human identity featured in the executive summary or key findings of a major breach report published in 2027.

Confidence: High.

5. Passkeys pass 45% of sign-ins

The prediction: The FIDO Alliance's next Passkey Index reports that passkeys account for more than 45% of successful sign-ins. Separately, at least one of Apple, Google or Meta makes new consumer accounts passwordless by default.

Why now: The FIDO Passkey Index 2026, published October 15, found passkeys carry 36% of successful sign-ins, up from 26% in 2025. Passkey sign-ins succeed 93% of the time, against 78% for traditional MFA. FIDO estimated in May that 5 billion passkeys are in use worldwide. Another ten-point jump would simply repeat this year's gain.

The second half has a precedent. Microsoft made new accounts "passwordless by default" in May 2025, so it does not count here.

How we'll know: the headline sign-in share in the first Passkey Index FIDO publishes after this one, if it appears by December 31, 2027. If FIDO publishes no index in 2027, that half is scored unresolved, not right. The second half needs a public announcement from Apple, Google or Meta.

Confidence: Medium. The 45% threshold is near trend. The passwordless-default half is the bolder call. It is a step toward the 2030 call on Future Tech's passwordless everything page.

6. Most EU countries still lack a working digital identity wallet at the end of 2027

The prediction: On December 31, 2027, fewer than 14 of the 27 EU member states offer a certified European Digital Identity Wallet that is generally available to the public, not a pilot. I also expect the December 24, 2027 deadline for businesses to accept the wallet to be delayed or narrowed.

Why now: EU law requires every member state to offer at least one certified wallet by December 24, 2026. In September, 24 of 27 states had yet to make a wallet available, according to Euronews data cited by Brussels Signal. Italy is the standout, with almost eight million monthly users. Germany set its launch for January 2, 2027. The Netherlands pushed its rollout to late 2027 after a pilot attracted 57 users. Privacy groups including EDRi also argue the implementing acts are diluting the law's safeguards.

How we'll know: the count of member states with a certified, publicly available wallet on December 31, 2027, and any formal change to the business acceptance date. The count is the primary test.

Confidence: Medium. The EU can move fast once a few large states launch, so this could go either way.

7. A card network writes dispute rules for AI agent purchases

The prediction: By the end of 2027, Visa, Mastercard or American Express publishes dispute or chargeback rules written specifically for purchases made by AI agents, such as a dedicated reason code for agent errors.

Why now: Agents already complete purchases on all three networks, but there is no clean way to separate an agent's mistake from a cardholder-authorized purchase. Worldpay put it plainly: "No liability shift exists yet." EMVCo's draft agentic payments framework took feedback until September 30, and Forkast reports it leaves non-fraud dispute liability unresolved.

Merchants absorb dispute costs by default today. Forkast cites PYMNTS Intelligence data that 93% of merchants believe AI or agent providers should bear the loss when an agent buys the wrong thing. That pressure will force a rule. Future Tech's agentic payments entry argues the deciding fight is over agent authorization, which this rule would formalize.

How we'll know: published network rules that reference agent-initiated transactions in the dispute process.

Confidence: Medium.

8. Post-quantum questions move from federal contracts into everyday security reviews

The prediction: In 2027, post-quantum cryptography readiness questions become standard in commercial security questionnaires and enterprise RFPs, well beyond defense and federal buyers.

Why now: From January 1, 2027, all new acquisitions for US National Security Systems must comply with CNSA 2.0, the NSA's quantum-resistant algorithm suite, as SafeLogic summarizes the NSA timeline. That gate forces every vendor selling into those systems to have a post-quantum answer. Once vendors have one, enterprise buyers will ask for it too. "Harvest now, decrypt later" applies to any data that must stay confidential for a decade. My post-quantum migration guide for authentication covers what that answer should contain.

How we'll know: post-quantum questions in a widely used standard questionnaire (such as the Shared Assessments SIG or the CSA CAIQ), or in published procurement requirements from major enterprises outside government. This is the near-term step toward Future Tech's post-quantum rollout call for 2030.

Confidence: Medium.

9. CMMC certification returns in a narrower form

The prediction: The Department of War restarts third-party CMMC certification in 2027, but for a narrower set of contracts than originally planned. False Claims Act enforcement becomes the main driver of defense contractor compliance.

Why now: CMMC Phase 2 would have required third-party Level 2 certification from November 10, 2026. The Department of War's CIO, Kirsten Davies, paused it on July 13, according to a Pyramid Systems review updated November 2. A September 3 policy memo told contracting officers to remove third-party assessment requirements from contracts. As of November 2, the department's CMMC page says implementation "is paused in Phase 1," with no new date.

Meanwhile, the Justice Department settled False Claims Act cases over NIST SP 800-171 gaps on June 18 and September 1, the second with Honeywell Aerospace for about $2.04 million. In my view, self-attestation backed by legal liability is already doing the work certification was meant to do.

How we'll know: whether third-party certification is reinstated in 2027 and for which contracts, plus the number of cybersecurity-related False Claims Act resolutions announced in 2027.

Confidence: Low. Policy direction is genuinely uncertain here, which is why I am saying so.

10. AI engines become the first security analyst most buyers consult

The prediction: By the end of 2027, more security buyers will form their first vendor shortlist from AI engines than from analyst reports. Security vendors will also begin reporting "AI share of voice" to their boards alongside pipeline.

Why now: Forrester reports that 94% of business buyers already use AI in their buying process. GrackerAI, my company, published eight AI search visibility benchmarks this year covering 1.74 million AI answers about security products. In the benchmarks that report a maximum, no two engines shared more than 0.29 of their cited domains on a Jaccard scale. In the SIEM benchmark, vendors' own websites supplied 94.9% of ChatGPT's citations. When AI shapes the shortlist, AI visibility becomes a revenue metric.

How we'll know: a published buyer survey showing AI engines ahead of analyst firms as a first research source, and at least one public security company referencing AI visibility in earnings materials.

Confidence: High on buyer behavior. Medium on board reporting.

One prediction I'm not making

AGI will not arrive in 2027. Jensen Huang declared it had arrived in September. I argued in AGI has not arrived that the definitional games behind such claims serve fundraising more than science. Everything in this list assumes something more useful: AI that is very capable, increasingly autonomous, and still lacking the judgment to be trusted without identity, authorization and monitoring around it. That gap is where most of 2027's security work will happen.

The thread connecting all ten

Read together, one theme emerges: identity is becoming the control plane for everything autonomous.

Agents need identities. Their actions need authorization. Their purchases need proof of intent. Their monitoring needs to be independent of them. Human sign-in is moving to passkeys and digital wallets so the human end of the chain is trustworthy too.

For twenty years, identity was about logging people in. In 2027, it becomes about governing what acts on their behalf.

The scorecard promise

In December 2027 I will publish a scorecard grading each prediction as right, wrong or partially right, with the evidence. If you think I am wrong about any of these, tell me now, and I will include the strongest counterarguments in the scorecard.

Which of these ten do you think I will get wrong?

FAQ

What are the biggest identity security trends for 2027?

Identity is becoming the control plane for autonomous systems. Expect formal rules for identifying and authorizing AI agents, and non-human identities such as API keys and agent credentials becoming a leading breach risk. Passkeys keep replacing passwords, and EU digital identity wallets roll out more slowly than the law requires.

How widely are passkeys used in 2026?

The FIDO Alliance Passkey Index 2026, published October 15, 2026, found passkeys account for 36% of successful sign-ins, up from 26% in 2025. Passkey sign-ins succeed 93% of the time against 78% for traditional MFA. FIDO estimated in May 2026 that 5 billion passkeys are in use.

What happens with CNSA 2.0 in January 2027?

From January 1, 2027, new acquisitions for US National Security Systems must comply with CNSA 2.0, the NSA's suite of quantum-resistant algorithms. It is the first hard procurement gate in the US post-quantum transition, and it applies to new purchases rather than systems already deployed.

Will EU countries meet the digital identity wallet deadline?

Most will not. EU law requires every member state to offer a certified digital identity wallet by December 24, 2026. In September 2026, 24 of 27 member states had yet to make one available. Italy is the main exception, with almost eight million monthly users.

How will these predictions be graded?

Each prediction states a test that can be checked by December 31, 2027. In December 2027 I will publish a scorecard marking each one right, wrong or partially right, with the evidence, following the method of my earlier public prediction scoreboard.