The proposed definition
What SEDOS describes
Synthetic Emergency Denial of Service, or SEDOS, describes a potential attack in which fabricated emergency reports, enabled by synthetic media or automated communication, overwhelm the operational capacity needed for genuine emergencies. The object of concern is not merely a busy communications system. It is the availability of people, attention, vehicles and specialist support when someone actually needs help.
The proposed model connects a false claim to its consequences for emergency response. A communications service could remain technically available while staff spend scarce time assessing fabricated incidents. Equally, a period of unusually high demand can be entirely genuine. Volume, repetition or technological novelty alone cannot establish malicious intent.
This is a threat model and a research question, not a recognised incident classification or a claim of institutional adoption. The useful question is whether existing resilience arrangements remain adequate as the capabilities behind deceptive communications change. That question requires evidence about both malicious reporting and the harms caused by defensive mistakes.
Evidence notes and source limitationsRelated but different threats
Swatting, overload and the boundary of the term
Swatting is already a documented danger: a false report can provoke an emergency response against a targeted person or place. The FBI describes the diversion of resources and risks surrounding this practice. It does not need AI to be harmful. Calling every swatting incident SEDOS would obscure what is already known and make the new term less useful.
Telephony denial of service, often abbreviated TDoS, concerns disruption to communications availability. Fraud involves deception but need not target emergency response capacity. Ordinary overload may follow a real disaster, an infrastructure failure or several unrelated emergencies. These categories can overlap in consequences without sharing the same cause.
SEDOS proposes a particular lens: synthetic or automated false reporting as a means of denying operational capacity. An isolated AI voiced hoax does not by itself demonstrate that broader pattern. Reporting should distinguish the mechanism observed, the effect documented and the intent established, rather than using a dramatic label to bridge gaps in evidence.
Evidence notes and source limitationsTechnology convergence
The concern is a convergence of capabilities
The conceptual concern is the combination of synthetic speech, language generation, conversational interaction and automated communication. Each capability deserves separate scrutiny. A system that generates fluent language does not necessarily understand an unfolding event, and a convincing voice does not establish a convincing or successful deception.
At a class level, these technologies can change how fabricated information is presented and how much human effort its preparation requires. The research question is whether combinations alter the burden placed on emergency institutions. This page does not describe attack systems, practical routes, scripts, evasion methods or scaling arrangements.
NPR’s reporting provides evidence that AI generated voices have entered the swatting context. It does not measure the overall prevalence of that use or demonstrate a coordinated campaign overwhelming services. CISA’s discussion of AI and fraudulent calls makes the subject relevant to resilience planning, without establishing that every proposed future scenario is happening.
A responsible assessment separates capability from deployment, deployment from effectiveness, and effectiveness in an individual deception from system wide loss of capacity. Each transition needs additional evidence. The existence of a tool is not evidence that an emergency service has been defeated by it.
Evidence notes and source limitationsWhat the evidence establishes
Documented incidents, carefully bounded
Reuters reported on 5 November 2026 that a wave of bomb hoaxes affected nearly 30 flights across Europe. On 11 November, it reported that around 6,300 schools in Italy received bomb warnings, alongside warnings to schools in Scotland, Northern Ireland and the Republic of Ireland. The Italian figure should not be presented as a combined total for all those countries.
There is no established evidence that AI was involved in those European incidents. They demonstrate disruption associated with false threats, not an actual AI enabled SEDOS attack. They also do not supply a measurement of emergency response delay or a basis for inventing casualties, costs or motives.
Separately, NPR reported that a hoax involving Wheatland Union High School in Yuba County used an AI generated voice. That is relevant evidence of reported AI assisted deception in an emergency communications setting. It does not establish mass SEDOS. Survey responses discussed in that reporting should not be converted into a verified percentage of emergency calls or a prevalence estimate.
CISA and the FBI provide institutional context about fraudulent emergency reporting and its dangers. NIST provides technical context about synthetic content transparency. None of these sources validates the defensive proposals below. The distinction between a reported incident, an advisory and a research proposal should remain visible whenever this page is quoted.
Evidence notes and source limitationsThe human stakes
The person waiting for real help
The central public interest is not whether an artificial voice sounds impressive. It is whether a person facing a genuine emergency can still obtain timely assistance. The FBI’s warning about resource diversion explains why false reports matter even when no communications network fails.
Defensive systems can create another source of harm if they cause a real emergency to be discounted. An unusual account, a frightened caller, a speech disability or a language unfamiliar to staff must not become shorthand for deception. This is a design requirement for the proposals here, not a claim that a particular service currently follows such a practice.
Resilience therefore has two obligations: protect capacity against deception and protect access against excessive suspicion. A tool that identifies more hoaxes but delays more genuine emergencies cannot be judged successful by its classification score alone. The people absent from a dataset, including those unable to complete a conversation, must be part of an evaluation plan.
Evidence notes and source limitationsAI and human judgement
AI may help, but it cannot be the truth oracle
The key research thesis is that AI may need to help assess whether an emergency report is a hoax or a genuine call for help, regardless of how that report was created. Detecting synthetic speech answers a different question from establishing whether an incident is real.
A person using synthetic speech as assistive communication can be reporting a real emergency. A human speaking naturally can make a false report. A recording with clear provenance can contain an untrue claim. Conversely, missing provenance does not make a frightened person’s account false. These distinctions must survive every interface label and procurement specification.
Potentially useful AI functions include organising an evidence ledger, surfacing contradictions, translating information for review and identifying uncertain relationships between reports. These are proposals for decision support, not demonstrated benefits in this setting. Translation and summarisation can introduce errors, so original information and the limits of the transformation must remain available to authorised reviewers.
NIST’s treatment of transparency techniques supports a cautious boundary: provenance and detection are partial tools, not a comprehensive truth system. In this proposed approach, a model can offer a bounded observation, express uncertainty or abstain. It cannot make an unexplained rejection of an emergency report the default result of suspicion.
Evidence notes and source limitationsProposed defence / Not operational guidance
An evidence ledger, not a suspicion machine
The following is a proposed defensive research architecture, not an implemented product or a validated emergency workflow. Its organising principle is to preserve what is known, show what is uncertain and keep human responsibility visible. Technical novelty is secondary to timely assistance.
- Keep an evidence ledger. Separate the reported claim, source, time, transformations, corroboration and unresolved questions. Record corrections and model versions. A model summary should link back to its basis rather than silently replacing the original account.
- Seek independent corroboration. Distinguish a genuinely separate observation from repeated copies of one claim. Provenance can clarify an item’s history but cannot certify the event. Two websites repeating the same report are not two independent witnesses.
- Bound anomaly grouping. Similarity signals should have limited scope, reviewable reasons and expiry. A cluster may reflect a genuine shared emergency. Grouping must not automatically merge distinct incidents or downgrade a person’s request for help.
- Calibrate uncertainty and allow abstention. Express the specific question a model addressed and how its confidence was evaluated. Missing information should remain visible. Where evidence or evaluation is inadequate, the system should say it cannot assess rather than invent certainty.
- Keep verification human led and parallel. Authorised staff should own corroboration and escalation within their existing responsibilities. Additional checking should support the response, not become an automated holding stage that delays help. This page specifies no dispatch decisions or verification routes.
- Design for safe load and graceful failure. Evaluate offline capacity planning, fallback access and staff workload. If a model or supporting service becomes unavailable, emergency access should not depend on it. Failure modes belong in the design before any live reliance.
- Constrain data and protect access. Apply purpose limits, minimum necessary data, restricted permissions, retention rules and auditable access. Test security and accessibility together. No voice type, identity credential or device capability should become a prerequisite for assistance.
These layers would need evaluation as a whole. An accurate component can still create an unsafe system if its presentation encourages overconfidence, its records expose callers or its checks consume the attention needed for a real emergency.
An open research agenda
Seven concepts worth testing, not trusting yet
All seven concepts are unvalidated. They are proposals for investigation, not established solutions or deployments. The sources explain relevant problems and technical limits; they do not endorse these designs.
UNVALIDATED / RESEARCH CONCEPT 01Federated incident similarity
- Goal
- Notice potentially related reports across services without building a central pool of emergency audio.
- Possible mechanism
- A research system could compare tightly limited incident descriptors held by participating agencies, returning an uncertain similarity signal rather than a shared recording or automatic verdict.
- Failure, privacy and access concerns
- Descriptors can still expose people or locations. Similar reports can describe a genuine widespread disaster. Different languages, incomplete descriptions and local recording practices can distort comparisons.
- Validation needed
- Test linkage errors, information leakage and benefit to staff using consented or appropriately governed data. Require local control, deletion rules and a way to investigate a match without treating it as proof of fraud.
UNVALIDATED / RESEARCH CONCEPT 02Proof of capture, not proof of truth
- Goal
- Help staff understand what is known about a recording’s origin while preserving access for callers with no compatible device.
- Possible mechanism
- Optional capture attestations could record that a supported device produced a particular item at a particular stage. The ledger would keep that claim separate from the reported emergency.
- Failure, privacy and access concerns
- A genuine recording can contain a lie. A compromised device can mislead. Assistive speech, older phones and edited evidence may lack an attestation for entirely legitimate reasons.
- Validation needed
- Evaluate compromise scenarios, accessibility exclusions and whether staff overtrust a provenance badge. Missing provenance must remain an absence of information, never a reason to withhold assistance.
UNVALIDATED / RESEARCH CONCEPT 03A defender and challenger model pair
- Goal
- Expose weak reasoning before an automated suggestion shapes human judgement.
- Possible mechanism
- One model could summarise available evidence while a challenger searches for unsupported inferences, alternative explanations and reasons to abstain. Staff would see disagreements and the underlying sources.
- Failure, privacy and access concerns
- Two models can share training data, blind spots and vendor infrastructure. Agreement is not independent corroboration. A persuasive challenge can also distract staff or amplify an irrelevant detail.
- Validation needed
- Measure correlated failures, staff workload and missed emergencies against simpler tools. Require tests with both models wrong together, and never represent their agreement as two independent witnesses.
UNVALIDATED / RESEARCH CONCEPT 04An emergency capacity digital twin
- Goal
- Explore resilience choices without experimenting on people seeking emergency help.
- Possible mechanism
- An offline simulation could represent broad demand, staffing and resource constraints, comparing continuity arrangements under several hypothetical conditions. It would be a planning instrument, not a live dispatch controller.
- Failure, privacy and access concerns
- Historic records can hide unmet need. Simplified travel, staffing or incident assumptions can produce false confidence. Small communities may be identifiable even in apparently aggregated inputs.
- Validation needed
- Publish assumptions, sensitivity ranges and uncertainty. Compare predictions with held out operational periods under proper governance, involve frontline staff, and keep any live deployment outside the scope of the experiment.
UNVALIDATED / RESEARCH CONCEPT 05Selective disclosure credentials
- Goal
- Let a willing participant disclose a limited verified attribute without exposing a full identity record.
- Possible mechanism
- An optional credential might confirm a relevant organisational relationship while revealing little else. It would supplement a report rather than certify its truth or establish a preferred class of caller.
- Failure, privacy and access concerns
- Credentials can be stolen, outdated or coercive. People without documentation, network access or institutional membership must not face a lower standard of care. Issuer records can become tracking tools.
- Validation needed
- Test revocation, privacy leakage and access disparities with affected communities. Evaluate whether any informational gain survives a strict rule that a credential is never a gate to emergency assistance.
UNVALIDATED / RESEARCH CONCEPT 06A protected verification side channel
- Goal
- Give authorised staff a separate place to resolve uncertainties while the emergency response continues.
- Possible mechanism
- A controlled staff workspace could hold corroboration requests, accountable handovers and the evidence ledger. Access would be restricted to authorised personnel and independent of public reporting.
- Failure, privacy and access concerns
- An extra workspace can fragment attention or become a sensitive intelligence store. Incorrect contact information, compromised staff accounts or stale records can make apparent confirmation unreliable.
- Validation needed
- Test access control, audit trails, staff workload and continuity during outages. Demonstrate that parallel checks do not delay necessary action. This proposal does not publish contact routes or verification procedures.
UNVALIDATED / RESEARCH CONCEPT 07Privacy preserving coordination
- Goal
- Help agencies recognise a shared capacity problem while minimising disclosure about individual callers.
- Possible mechanism
- Research could compare coarse, purpose limited capacity signals and governed incident summaries. Shared information would have an explicit operational purpose, retention period and accountable recipient.
- Failure, privacy and access concerns
- Even aggregate signals can identify a person in a sparsely populated area. Combining datasets can reveal more than any contributor intended. Shared errors can spread rapidly across institutions.
- Validation needed
- Assess reidentification, malicious participants and the utility of less detailed alternatives. Require legal review, clear correction routes and tests showing that coordination improves resilience without creating a permanent surveillance layer.
Testing and measurement
Measure missed emergencies and delay first
Independent evaluation should begin with the consequences that matter: genuine emergencies missed, time added before appropriate assistance, staff workload and disparities between caller groups. False alarms, unnecessary escalations and the burden of correcting a model also matter. A single headline accuracy figure conceals these tradeoffs.
Consider a purely illustrative dataset of 10,000 reports, of which 100 are hoaxes. A hypothetical detector that flags 90 of those hoaxes and incorrectly flags 1 per cent of the 9,900 genuine reports produces 189 flags. Of those, 99 concern genuine reports. Fewer than half of the flagged reports are hoaxes. These invented numbers illustrate base rate effects; they describe no service, vendor or observed incident.
Testing should include diverse languages, accents, speech disabilities, assistive speech, distress, trauma and limited connectivity. Performance should be reported separately where sample sizes permit meaningful analysis, with uncertainty and data limitations disclosed. A missing evaluation group is a gap to investigate, not evidence of equal performance.
Before staff rely on a system, a governed silent trial could compare its suggestions with existing practice without allowing those suggestions to change live decisions. Such a trial still processes sensitive data and needs permission, safeguards and oversight. Independent reviewers should examine disagreements and delayed or missed emergencies, rather than assuming the historical record is perfect ground truth.
Later stages would need explicit stop criteria, frontline participation and a controlled evaluation design. Evidence from a vendor demonstration or a convenient test set is not enough to establish safety in a different jurisdiction. No effectiveness claim is made for the concepts on this page.
Evidence notes and source limitationsContinuity and safe load planning
Plan for demand without rehearsing harm
The proposed planning emphasis is continuity under uncertainty. Services could evaluate staffing resilience, accountable handovers, accessible fallback arrangements and the availability of specialist support when information is incomplete. This is an agenda for authorised planning teams, not a dispatcher protocol or a description of any agency’s procedures.
Exercises should be offline or otherwise isolated from live emergency access. They should not generate deceptive reports into operational systems or consume capacity reserved for the public. A simulation should include genuine surges, infrastructure faults and uncertain information, not only malicious scenarios that make a particular product look useful.
Useful outputs would include documented assumptions, capacity limits, dependencies and unanswered questions. A planning model cannot promise that resources will be available in an actual event. The test is whether an organisation understands its options and can maintain service without turning an automated suspicion score into a barrier to care.
Evidence notes and source limitationsProcurement and accountability
Buy evidence, preserve a way back
A procurement evaluation should ask what task a product actually performs, what evidence supports it and what remains outside scope. Require independently examinable test results, representative evaluation data, subgroup analysis and clear documentation of uncertainty. A supplier should distinguish synthetic content detection from assessment of an incident’s truth.
Proposed contractual requirements include audit access, incident disclosure, strict limits on secondary use of emergency data and an explicit prohibition on training unrelated products with that data without an appropriate lawful basis and governance. Security review should cover access control, dependencies, retention, deletion and recovery from compromise.
Every material model, threshold or interface change should have a recorded owner, evaluation and approval process. Version history should make past suggestions reconstructable. A rollback plan should restore a known safe operating arrangement when performance, security or access deteriorates; removing automation must not remove the underlying ability to receive help.
Public accountability should describe purpose, safeguards and evaluation findings without exposing personal information or sensitive operational arrangements. Affected communities, accessibility specialists and frontline professionals should help define acceptable use. Corrections and complaints need accountable human review. These are proposed governance conditions, not claims that a regulator has certified this architecture.
Evidence notes and source limitationsEvidence desk
Sources, scope and further reading
Read each source for the claim it supports. Reporting, institutional advice and technical reviews answer different questions. External source links open in a new tab. PDF documents are identified below.
- New wave of bomb hoaxes hits dozens of flights across Europe ↗Reuters · 5 November 2026
Reports disruption affecting nearly 30 flights across Europe. Evidence of disruption from false threats, not evidence that AI was involved.
- Hoax bomb alerts sent to schools in Italy, UK and Ireland in latest disruption ↗Reuters · 11 November 2026
Reports warnings received by around 6,300 schools in Italy, alongside warnings to schools in Scotland, Northern Ireland and the Republic of Ireland. Does not establish AI involvement.
- Swatters are using AI voices to deceive 911 dispatchers, and it’s only getting worse ↗NPR · 11 November 2026
Reports that a hoax involving Wheatland Union High School in Yuba County used an AI generated voice. This is reporting about AI assisted swatting, not proof of a mass SEDOS campaign. Survey responses mentioned in the article are not a measured prevalence estimate.
- An Introduction to Swatting and Fraudulent Calls for Emergency Communications Center Administrators ↗CISA · Emergency communications guidance · PDF
Discusses fraudulent calls, including AI, and technical and operational overload. This administrative resource provides context, not validation of the research proposals on this page.
- Threat Actors Use Swatting to Target Victims Nationwide ↗FBI · 29 April 2025
Explains swatting and the dangerous diversion of emergency resources. Supports the distinction between an existing harmful practice and the proposed SEDOS threat model.
- Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency ↗NIST · November 2024 · PDF
Reviews provenance and detection approaches, their limitations and evaluation needs. Provenance does not establish that a claim is true, and no single technical approach is a comprehensive solution.
- GDELT DOC 2.0 API Debuts ↗GDELT Project · API documentation
Documents translated news search covering 65 languages. That capability does not establish complete incident coverage, equal coverage across languages or independent verification of news reports.