
Cyber Risk from Early Hacking to Agentic AI
Quantum computing and frontier scenarios
Quantum risk is a transition, not a countdown
FIELD NOTE Evidence date: November 19, 2026 · Protect data by its required secrecy lifetime, not a headline estimate
What is already actionable
The principal public-key risk is conditional on future cryptanalytic hardware. But cryptography does not change instantly across endpoints, software, hardware and vendors. Financial, government and industrial systems can retain sensitive material past the lifetime of the products that first handled it. NIST finalized its transition guidance on October 27, 2026, after approving three initial post-quantum standards on August 13, 2024. This is a concrete modernization program, not a calculation of the date on which an adversary will possess the necessary machine. [129] [130]
The right question is not “will quantum computers threaten us tomorrow?” It is whether a service or record will still need protection during the possible interval between the expansion of future capabilities and the completion of a well-tested migration.
Four developments should not be blended
Engineering progress in quantum computing is distinct from cryptographic standards and actual deployment. A report of a laboratory milestone can change confidence about one engineering constraint but does not establish that communications protected by currently deployed algorithms have been broken. An implementation offering a standardized quantum-resistant algorithm does not prove it is deployed across a whole system. [141] [129]
Likewise, AI agents run on today's classical infrastructure, and their permissions and reliability drive risk now. AGI refers to a contested broader capability claim. Quantum acceleration of a narrowly defined task does not imply AGI, and AGI would not automatically deliver the fabrication, error correction and real-time operations required for a cryptanalytically relevant quantum computer. Prepare for these axes separately rather than waiting for a single convergence event. [147] [131]
Confidentiality and signatures fail differently
FIELD NOTE A channel can use one algorithm to agree on a key and another to authenticate the server
Secrecy
Public-key key establishment allows parties to agree on a shared secret over an untrusted network. Modern systems often use that shared secret with symmetric cryptography to protect actual message traffic. A future quantum computer with sufficient scale and error correction would threaten the mathematical assumptions behind widely used traditional public-key systems. Changing the cipher used for bulk data alone does not necessarily address an exposed public-key establishment mechanism. [129]
This distinction answers why a familiar service can be at risk while still using a strong approved symmetric cipher: its session key, storage key or root of trust may depend on something else.
Integrity and authenticity
Digital signatures protect identity and integrity, including for signed software releases, trust anchors and authenticated documents. A future capability against existing signing keys could undermine confidence in who approved an artifact; that is distinct from reading historical communications. A material leak today and a trust-chain failure tomorrow call for different contingency planning. [130] [129]
Do not infer that all symmetric encryption, hashes or key-management functions need the same replacement as public-key mechanisms. NIST does not expect to require a wholesale move away from current symmetric standards for this migration. In practice, validate both the primitives and the overall trust design. A secure algorithm cannot repair stale keys, undisclosed recovery paths, a corrupted policy store or a rollout that omits a dependency. [129] [138]
Harvest risk starts before a new machine exists
FIELD NOTE Scope defense by retention, authorized access and data in transit
Exposure can begin during collection
The National Cybersecurity Center of Excellence distinguishes harvest-now, decrypt-later risk: sensitive encrypted information retained today may still be valuable when future cryptanalytic capabilities exist. Whether this materializes depends on what was collected, the original protections, the time value of confidentiality and the resources an adversary eventually has. Migration cannot retroactively replace protections on copies already outside an organization's control. [131] [129]
Inventory classes whose sensitivity outlasts routine technology refresh. Examples include health records, contractual archives, trade secrets and protected identities. A long business retention period is not necessarily a long confidentiality period, so obtain the data owner's assessment.
A defensible starting order
Start with long-lived sensitive data in transit, then check storage and backup key management as applicable. Record the route actually used by a record and whether external services terminate and recreate communications protection. Minimize unnecessary copies and retention in line with legitimate obligations. Retain organizational context for legal records and avoid deleting evidence required by policy. [139] [129]
For a fictional service whose records need ten more years of confidentiality and whose migration could take four, “we can wait for clearer hardware evidence” is not necessarily a defensible answer. Those numbers are planning inputs, not a quantum-arrival estimate. Elevate cases where lengthy migrations, irreversible disclosure or weak control over custody overlap, and reassess annually as timelines and asset ownership change. [131]
The first three standards solve different problems
FIELD NOTE Key establishment: FIPS 203 · Digital signatures: FIPS 204 and 205
What was approved
FIPS 203 specifies ML-KEM, a key-encapsulation mechanism that helps establish a shared secret. FIPS 204 specifies the module-lattice-based signature scheme ML-DSA. FIPS 205 specifies the stateless hash-based signature scheme SLH-DSA. These are algorithms for conventional computing infrastructure; an organization does not need quantum hardware to adopt them. A signature scheme cannot substitute for a key-establishment mechanism, or vice versa. [130] [131]
Cryptographic diversity provides options across deployment constraints and helps manage dependency on a single mathematical family. It does not eliminate the need for rigorous verification of the implementation and its integration.
A candidate is not a product certification
NIST selected HQC as an additional key-encapsulation option in 2025. The October 2026 transition guidance still describes an HQC standard as under development and separately discusses additional signature work. Selection, standard publication, protocol integration, module validation and secure deployment are independent milestones. [140] [129]
Organizations should ask which service functions a vendor covers, which versions and supported platforms are in scope, and what evidence shows security is active rather than simply available. Prefer vetted library and protocol support appropriate to policy; do not design a custom construction or treat an algorithm name as an assurance statement. Document approval and compatibility for transit protection and software signatures separately. [138] [139]
Federal deadlines do not forecast quantum hardware
FIELD NOTE U.S. dates cover different instruments, system scopes and control types
Read the final timeline
Under June 2026 Executive Order 14412 and OMB M-26-15, the most sensitive federal systems follow a faster track. NIST IR 8547 gives December 31, 2030 for key establishment and December 31, 2031 for digital signatures on federal high-value assets and high-impact systems, and describes prioritizing comparable high-sensitivity systems. OMB's memorandum does not apply to national security systems, which have distinct policies and should be assessed against applicable guidance. [132] [133] [129] [134]
NIST's broad trajectory is deprecation of vulnerable public-key mechanisms after 2030 and disallowance under its standards and guidance after 2035. These are not predictions for a machine or blanket deadlines for every firm.
Avoid three calendar errors
First, federal support for TLS 1.3 by January 2, 2030 is a platform step, not proof that post-quantum key establishment or signatures are in use. Second, do not use the general 2035 transition as permission to defer high-priority assets with earlier applicable dates. Third, an order directing work on a contractor rule does not establish that the proposed requirements are already final or apply to a particular contract. Check actual clauses and subsequent rules with qualified advice. [132] [133] [129]
For multi-market organizations publish an applicability table: authority, asset class, required behavior, deadline, evidence owner and latest source. When dates differ, use the applicable schedule and consider taking reversible preparatory action early. [138]
Exhibit 12 · PQC deadlines are not quantum arrival dates
NIST approved FIPS 203, 204 and 205 in August 2024. NIST summarizes accelerated federal deadlines for certain high-priority systems and a 2035 transition target, with general NIST disallowance after 2035. The dates describe standards and federal planning obligations, not predictions of a cryptographically relevant machine.
NIST IR 8547 final, October 2026: primary source
National roadmaps should be compared by scope
FIELD NOTE Dates: U.K. 2028 / 2031 / 2035 · Canadian federal non-classified roadmap: 2026 / 2031 / 2035 · EU coordination: high-risk by end-2030
The United Kingdom and Canada
The UK NCSC targets initial discovery and a migration plan by 2028, highest-priority migration by 2031, and completion across systems, services and products by 2035. The Canadian Centre for Cyber Security's roadmap for federal non-classified systems sets an April 2026 initial departmental plan and progress reporting, high-priority completion by end 2031 and remaining systems by end 2035. These sources use different institutional scopes and “completion” can depend on how vulnerable functions are handled. [135] [136]
A side-by-side date chart can inform staffing and procurement, but neither comparison nor coincident dates establish mutual recognition, identical control definitions or uniform enforcement.
European Union and supplier sequencing
The EU coordinated roadmap asks for a phased start and prioritizes high-risk cases no later than the end of 2030. Its status should not be confused with every member country's applicable requirements or laws. Check sector obligations, cloud locality and contractual roles before assuming that one baseline controls all assets. [137]
An international network can have cryptography enforced by software developers, cloud providers, device vendors and customer-controlled trust operations. Record which jurisdiction's guidance applies to a service and which party can deploy each change. For a partner connection, ask when both sides can support interoperable transport and what happens while only one side has transitioned. Include the older endpoint population separately, instead of letting an average hide it. [138] [139]
Physical qubits are not logical computing capacity
FIELD NOTE Error correction turns a hardware demonstration into evidence only for the scope tested
The pivotal measurement
A physical qubit is a quantum hardware element. A logical qubit is encoded so that a larger system can protect quantum information against errors. The relationship is not fixed: physical quality, architecture, encoding, control errors, measurement, decoding and the intended workload change overhead. In the [*REDACTED*] below-threshold memory study, a 101-physical-qubit configuration encoded a distance-seven logical memory with 0.143% ± 0.003 percentage points error per correction cycle. The memory result was published online in December 2024. [141]
That is not 101 usable logical qubits; nor does the reported error rate describe every gate, scaling behavior or workload on another device.
Broader constraints
A controlled memory demonstration is still meaningful. Improving the error rate as the tested code size grows supports the prospect of scaling. But an operational computation needs multiple logical operations, sustained error control, long runs, clean orchestration and usable throughput. Hardware packaging, calibration, control electronics and manufacturing can impede repeatability and scale. The published study also reports rare correlated errors, a reason not to extrapolate a typical error rate mechanically. [141]
Read milestone reports for the tested logical task, wall-clock run duration, error metric and comparison baseline. Device counts from incompatible architectures do not by themselves rank their practical capabilities. Watch for separately verified demonstrations covering whole workloads under realistic operating conditions. [131]
Quantum benchmarks answer narrow questions
FIELD NOTE A result can establish laboratory progress without proving a new general capability
A benchmark is a defined problem
[*REDACTED*] described [*REDACTED*] in October 2025 as a demonstration on a set of benchmarking circuits measuring a quantum observable. Reproducibility across relevant systems and comparisons with known classical methods make the result interesting. It does not establish that the computer accelerates arbitrary software, that the advantage extends to all inputs, or that it can break any real cryptographic protections. A revised classical baseline may alter the comparative claim without invalidating the underlying physical experiment. [144]
In October 2026, [*REDACTED*] also reported error improvements in a [*REDACTED*] experiment addressing a component of fault-tolerant quantum computation. Strong evidence for the component is not completion of the full machine. [142]
Evidence categories for procurement
Separate theoretical complexity results, simulations, calibrated laboratory observations, validated pilots and end-to-end commercially useful performance. Ask if resource claims refer to logical or physical units, if preprocessing and readout are included, and whether competing methods use comparable quality targets and up-to-date hardware. Avoid treating sponsor roadmaps and teaser demonstrations as independent corroboration. [141] [144] [145]
In a defensive monitoring program, update technical scenarios when the evidence changes enough to alter a real dependency. Usable signals include error correction across larger systems, reliable end-to-end computation, clearer reproducibility or better vendor support for migration. Hardware skepticism is not a reason to defer inventory; that work is useful even if expectations move later. [138]
AI can assist quantum research without creating AGI
FIELD NOTE Research support runs largely on classical systems; a quantum benchmark is not a model-performance benchmark
Classical AI for a quantum task
[*REDACTED*] is a neural decoder trained to infer error information for quantum error correction. The research tests a specific bottleneck using simulated and experimental data. Better decoding may help preserve encoded quantum information, but this does not remove fabrication and scaling constraints or automatically extend to lower latency or future architectures. A neural network assisting a quantum device is not the same as a neural network being trained on a quantum computer. [143] [141]
AI may also help analyze experiments or prioritize design alternatives. Demonstrate the benefit experimentally, count engineering time and failures, and assess how well it transfers.
Quantum computation for AI tasks
Proposed quantum applications for learning include specialized classification, dimension reduction and physical simulation. A 2026 study demonstrates theoretical results under particular access and comparison assumptions and gives application examples. That is not a general inference speedup or proof that fully operational devices with the claimed logical capacity exist. Require accounting for data access, error correction, input/output, comparison baselines and cost before carrying a specialized result into AI production plans. [145]
AGI is a separate and contested claim about breadth and adaptability. Neither a quantum experiment nor a benchmark on one model assistance task establishes general autonomy, open-ended competence or reliable control. These technologies can develop in parallel without converging into a single event. [147] [148]
Singularity is a scenario, not a dated event
FIELD NOTE Differentiate assistance, autonomous work, feedback acceleration and loss of effective control
Translate the label into testable claims
The term “singularity” is used for an hypothesized rapid transformation in capabilities, sometimes connected to AI systems improving the research processes that produce later systems. It is not an engineering milestone with a single accepted definition. Break it into narrower observable claims: does AI assistance raise quality-adjusted research output, across which tasks, with what level of human work, and over what period? Then ask whether faster outputs persist through data, validation, compute and deployment bottlenecks. [149] [148]
Automation in code or experiments may speed useful pieces of a pipeline without automation of the entire pipeline.
Distinguish acceleration from loss of control
METR's November 2026 assessment of one company's October research and development found evidence consistent with meaningful assistance under measured conditions, and identified reasons its estimates could be too high or too low. Internal self-reports, model access and benchmark selection do not establish a counterfactual economy-wide rate or autonomy without human oversight. [148]
The International AI Safety Report distinguishes relevant precursors from the sustained bundle of capability and access contemplated by loss-of-control scenarios. Experts disagree about likelihood and pathways. A defensible control agenda does not require a precise probability: preserve independent authorization, the ability to revoke capability and access, observations that could invalidate assumptions, and a tested ability to pause and recover. [147]
A convergence scenario needs independent conditions
FIELD NOTE The future threat changes only when capability, resources, deployment and trust failures coincide
Scenario: faster research, delayed hardware
A capable tool speeds experimental analysis and design, but equipment availability and dependable error correction still constrain the rate of hardware scaling. More measurements are possible; the usable workload does not immediately expand. Under this scenario the organization should continue standard migration on schedule, evaluate improvements for applicability, and not weaken ordinary identity and access controls in anticipation of hypothetical quantum activity. [143] [141] [129]
Evidence changing the scenario would include repeated, independently corroborated end-to-end engineering gains rather than new claims based on the same benchmark.
Scenario: financial and institutional controls break first
An agent may acquire broad authority because teams want to save time, long before a quantum device can materially affect production cryptography. If logs are absent, supervisors lack time to review actions, or credentials cannot be revoked independently, better mathematical protections do not cure the governance failure. A simultaneous pressure to update cryptography can amplify operational load and make change control more difficult. [147] [138]
Set separate triggers for model permissions and cryptographic readiness. Rehearse stopping automated work safely, preserving logs and restoring known authorized behavior. Require authorized people with tested access to decide the next step. Avoid tying an organization's readiness to an AGI declaration, a prediction market, a speculative hardware date or an unverified provider headline. [147] [131]
Build a cryptographic inventory that can be verified
FIELD NOTE Know every covered service and evidence owner, not merely every vendor
Identify actual uses
List applications and environments, cryptographic function, protocols, algorithms, key and certificate ownership, data lifetime, vendor dependencies, hosting boundary, hardware limitations and available telemetry. Prioritize externally exposed flows, high-value data and long-lived verification roots. Keep storage, transit, signing and recovery paths separate. Procurement claims and product configurations are inputs, not proof production is using a particular mechanism. [139] [129]
Where discovery is incomplete, state which inventory method was used and where manual or vendor confirmation is outstanding. Publish confidence bands or counts by evidence stage rather than an estate-wide percentage based on assumed coverage.
Assign an owner to every unresolved item
A dependable record has an accountable service owner, root cause of the barrier, latest release or replacement milestone, and a mitigation or exception expiration. Include dependency owners who can change gateways and certificate chains, not simply the team paying the invoice. Track externally managed services and cryptographic hardware and firmware that may require replacement. [138] [135]
Validate samples of production negotiations where authorized; an advertised feature can coexist with unprotected segments. A provider's September 2026 visibility tooling is an example of distinguishing client-facing and backend protection in practice, not evidence for every network or any particular organization's configuration. [146]
Build crypto agility without creating fragile fallback
FIELD NOTE Three measures: supported function, observed use, tested recovery
Flexibility is a security property
Crypto agility means the ability to replace or adapt cryptographic mechanisms without sacrificing security or continuity. It includes protocol options, configuration ownership, centrally maintained software dependencies, replacement paths for firmware, certificate lifecycle readiness and policy enforcement. More settings are not automatically more agility: unowned or unrestricted switching can create ambiguity and increase reliance on obsolete defaults. [138]
Select supported configurations and test them in owned environments. Establish a written decision rule for failure, a boundary on fallback and a process for retiring vulnerable paths. Do not extend exceptions indefinitely because an obsolete dependency is expensive.
The transition period is measurable
Run a narrowly scoped pilot before broad adoption. Compare service success, failure modes, latency, resource use and backward compatibility with prior configuration. Include infrastructure limits and device populations; key, message and signature sizes are deployment inputs, not proof of operational performance. Confirm trust path behavior, key rotation, monitoring and interruption recovery through authorized exercises. [139] [129]
Measure agreed denominators: migrated high-priority connections as a fraction of all enumerated high-priority connections; tests passed by normalized environment; exception count, severity and age. Identify unobservable routes; a green score from a partial population can be misleading. Require official protocol and vendor guidance, and document why a temporary exception is bounded and approved. [138] [131]
Signing and archival trust need their own plan
FIELD NOTE Firmware and regulated records can live beyond the issuing service
Identify who will verify older artifacts
A signed release, document or firmware package may be checked long after it was issued. Record verifiers, future trust material, reissuance options, existing timestamp evidence and the consequences of rejecting an old object. Distinguish ability to validate a legacy object from authority to produce a new one. Note whether older field devices can accept new trust material and whether a reachable maintenance process exists. [129] [138]
Prioritize long-lived assets where delayed migration is expensive or where loss of a trust anchor would shut down a critical service. Set compatibility bounds so a transition does not silently preserve outdated protection indefinitely.
Make recovery independent of the failed path
Consider a fictional fleet that can receive new signing keys only through the update process the organization wishes to replace. The recovery question is not merely whether new signatures exist, but whether the verifier can acquire a trustworthy new root if the old process must be retired. Plan and test a vendor-supported, authorized transition path before a cutoff. Maintain protected backups of trust configuration and rehearse replacement and validation. [135] [138]
Archive validation may involve durable provenance, timestamps, or re-attestation where suitable. Requirements depend on law, contractual obligations, and industry policy. Avoid revoking all old credentials, discarding evidence, or treating ordinary operational re-signing as a substitute for specialist legal review. [129] [131]
Prepare a rebound plan for failed migrations
FIELD NOTE If a control rollout causes disruption, recover within an approved protection boundary
What failure can look like
A phased rollout can disrupt a partner connection, device enrollment or certificate-verification chain if dependencies behave differently in production. This can be operationally serious without being a breach. Document the symptom, which trust functions are affected, and whether confidentiality or integrity was actually impaired. Preserve telemetry and bring the service owner and incident lead into the same decision process. [139] [138]
Maintain a roster of who can stop a rollout, where configuration evidence lives and which service continuity obligations constrain the response. Logs should capture actual controls in use without revealing secrets.
Test the rebound before the emergency
For approved test changes, define an authorized rollback and expected outcome. If any provisional fallback is needed, evaluate its exposure, scope, end date and approval requirements; do not normalize continuing on unsafe paths. Identify priority services that should isolate a route rather than silently drop safeguards. Capture observations before and after remediation so success means both restored operations and restored protection. [138] [131]
Measure time to detection, verified restoration, restoration of the intended guarded state, and number of exceptions left open. Update runbooks and dependency inventories after the exercise. Recovery that recreates obsolete trust across more of the system is not a rebound. [138]
Known, disputed and unknown claims need different action
FIELD NOTE Do not use the unknown as a reason for paralysis or as evidence for a claim
Well evidenced and actionable
NIST has approved three principal standards, issued final transition timelines and supplied discovery and agility guidance. Public experimental work has shown scoped improvement in quantum error correction. Public reports identify continued uncertainty about autonomous AI systems and loss of control. These facts can justify defensive preparation now. None supplies a precise date for a general cryptanalytic machine, a probability of disaster or evidence that an operational attack succeeded. [130] [129] [141] [147]
When analysis references a provider claim, include its benchmark, denominator, date and what corroboration is available. Preserve contrary results and changes to methods.
Open questions and leading indicators
Watch incomplete inventories, hard-to-replace roots of trust, costs across real deployment populations, changes in error-correction scale, and independent replication. For convergence, track measured human workload after AI assistance, the duration and authority under which agents act, and whether containment still works after a vendor or service failure. The largest unknown may be not raw capability but whether organizations deploy it with insufficient independent oversight. [138] [148] [147]
Escalate on meaningful evidence: overdue vendor commitments for critical systems, confirmed breaks in coverage, loss of auditability or verified claims that change a tested assumption. Avoid forecasting by combining unrelated trend lines. Publish what evidence would cause a risk rating or migration priority to change, with a date and owner. [131]
A practical sequence for the next twelve months
FIELD NOTE Suggested operating cadence; follow applicable directives and approved controls
First 90 days: discover and assign
Assign a cryptography lead, an AI deployment owner and an incident decision-maker. Create the initial inventory and prioritize long-lived sensitive information, externally reachable services and verification roots. Collect vendor roadmaps and note where primary evidence is absent. Select one high-priority service and one low-risk pilot that tests readiness without exposing sensitive material. Track actual use alongside available support. [139] [135]
For systems using AI agents, document permission boundaries, human approvals and trusted logs. Ask whether a non-agent process can revoke authority and restore known behavior if the system or its provider fails. [147]
Next quarters: test, assess and fund
Use approved standards and authorized environments to pilot changes; verify relevant parties, including earlier devices and recovery processes. Separate key establishment and signatures on the dashboard. Define expected migration stages, safe recovery, exceptions and which milestones warrant new funding. Communicate with vendors and partners before their own replacement windows close. [129] [138]
Complete an executive review that reports actual coverage, unmeasured populations, unresolved dependencies and validated recovery. Reassess against new official guidance and independently evaluated milestones rather than a forecast for a singularity event. For small operators, a verified vendor path plus an owner, dates and tests can be more useful than a sophisticated risk score with unsupported inputs. [131]
Exhibit 13 · Post-quantum migration is a staged program
Start with cryptographic and dependency inventory, prioritize sensitive long-lived information and important services, then pilot and phase migration. Track compatibility, vendor support, rollback and residual legacy exposure. Milestones are operational, not forecasts of future hardware capability.
48PEGASUS Research conceptual exhibit. Not empirical data.
Browse: White Papers · Reports · Downloads
