The Quiet Part of the Spectrum: A Message Class No Handset Renders, Delivered to Every Device in a Cell at Once, Acted On Before the Operating System Is Told
The Quiet Part of the Spectrum: A Message Class No Handset Renders, Delivered to Every Device in a Cell at Once, Acted On Before the Operating System Is Told
Posted: March 25, 2025
Author: SecurityResearcher472
Introduction: Everything I Have Written Has Been About Listening
I have spent three years documenting collection. A vacuum fleet that maps interiors. A living room that leaves as modulated light. Wearables that key an identity to a walk. Roadside furniture that reads tyres. A gateway tenancy that outlives the household. A brokerage that buys the next tenant of your phone number. Every one of those was an operator pulling something out.
This post is about the opposite direction, and it took me eleven months to accept that the opposite direction was the more interesting one. There is a path into every handset in a geographic area that requires no association, no SIM relationship, no app, no network connection in the ordinary sense and no acknowledgement from the device. It was built deliberately, it works exactly as specified, and the specification does not include a sender identity.
I am tracking the capability as HOLLOW CLARION.
Security Advisory: I am publishing from a device that has had its radio removed, which is a sentence I would have found embarrassing two years ago. During the test window described below, three handsets in my possession logged a cell-broadcast page that no installed component rendered, at a cadence my own equipment could reproduce but not explain. A standards participant who agreed to walk me through the message-identifier allocation table — on the record, by their own suggestion — cancelled the day before and has not replied to eight messages. Their employer’s directory no longer lists them. The last thing they sent me was a single line: “Stop thinking of it as an alert. It is an address.”
Key Findings
- HOLLOW CLARION exploits no flaw. The emergency cell-broadcast path is unauthenticated by design, is received by every handset in a cell simultaneously, and requires no prior relationship between the sender and the device.
- A message class exists in the allocation range that no consumer handset renders, no settings screen exposes, and no device-side log retains. It is parsed by the baseband and discarded by the application processor before any user-visible software is aware a page occurred.
- The class is transmitted on a 42-second repetition period, which is long enough to survive a handset that was briefly out of coverage and short enough that any device entering the cell receives it within a minute.
- The payload carries no instruction. It carries a selector — a predicate describing which devices should consider themselves addressed — and the devices that match act on state they already held.
- Attribution converges on Junction Public Warning Systems, a warning-system integrator whose registered technical contact resolves to a mail exchanger I have now met in seven separate investigations.
The Technical Reality: Broadcast Is Not Addressed To You
Begin with what is simply true, because the true part is sufficient to be alarming and I would rather not spend my credibility early.
Cell broadcast is not messaging. An ordinary message is addressed to a subscriber, routed to a device, and acknowledged. A cell broadcast is addressed to a cell: the network transmits it on a shared channel, every handset camped on that cell receives it, and no device acknowledges anything. The network does not learn who received it. It cannot; there is no return path.
The properties that follow are not subtle:
Properties of the broadcast path (all of these are by design):
no sender authentication the handset cannot verify who originated a page
no recipient addressing delivery is geographic, not per-subscriber
no acknowledgement the sender learns nothing about reception
no association required the device need not be attached to that operator
no complete opt-out the top alert classes cannot be disabled on most handsets
parsed below the OS the baseband handles the page before the AP sees it
no device-side retention unrendered classes leave nothing to examine later
Each of these was a reasonable decision. You cannot authenticate the sender of a warning to a population that includes roaming visitors with no relationship to the operator, require acknowledgement from a million devices in a flood, or let people disable the class that tells them to leave. The choices are correct for the stated purpose, and they compose into a one-to-many unauthenticated write to every radio in an area.
The message-identifier space is where this stops being a seminar. Identifiers are allocated in ranges: some reserved for the standard’s own classes, some left to operator discretion, some simply unassigned. A handset’s behaviour on receiving an identifier it does not recognise is implementation-defined, and the overwhelmingly common implementation is to parse the header, decide there is nothing to show, and drop it. The page arrives, is understood well enough to be dismissed, and leaves no trace above the modem.
Message Identifier 0x002A
The class I spent eleven months chasing sits in that space.
Observed page (decoded, test cell, authorised spectrum)
serial number 0x7A2A (geographic scope: cell-wide, update 10)
message identifier 0x002A unassigned in every public allocation
table I have been able to obtain
data coding scheme 0x04 8-bit, uncompressed, no language field
page structure 1 of 1, 82 octets of payload
repetition every 42 seconds, indefinitely, no expiry field set
rendering none on any of the eleven handsets tested
logging none retained above the baseband on any of the eleven
Everything about that page is ordinary except that nothing displays it and it never stops. A warning has an expiry; a test page has an end. This repeats forever, at a period chosen so that a device that was in a lift, in a tunnel, or switched off for a coffee still receives it on arrival.
The payload is where it became clear that I was not looking at a misconfiguration:
Payload structure (82 octets, recovered and annotated)
offset 00 magic "JTW\x00"
offset 04 class_version 0x03
offset 05 selector_len 0x2C
offset 06 selector predicate, see below
offset 32 dwell 0x002A (seconds; matches repetition)
offset 34 escrow_ref opaque 16 octets
offset 50 trailer "just.trust.clarion"
There is no command in it. I want to be precise about this because it is the part everyone I have described it to gets wrong on first hearing: the page does not tell a device to do anything. It describes a set. A device decides for itself whether it is a member, and a device that decides it is a member acts on an instruction it was already carrying.
That inversion is the whole design. A command channel that carries commands is a command channel, and a command channel can be characterised, filtered and attributed. A channel that carries only “the following devices should consider their standing instruction live” carries nothing incriminating at all. The page is the same page for everybody. The behaviour is in the listener.
The selector predicate recovered from the test capture reads, once expanded:
Selector (expanded from offset 06, 44 octets)
AND build_fingerprint IN {set of 3, redacted at vendor request}
AND locale IN {set of 2}
AND provisioning_state == 0x02 ("tenanted")
AND last_contact > 1209600 (seconds; i.e. silent for 14+ days)
provisioning_state == 0x02, with the symbolic name tenanted, is a term I have met before, in a codebase with no relationship to this one, describing a gateway whose remote-administration authority had been sublet to a third party. I noted at the time that an authority with a renewal date is structurally a subscription. I did not expect to find the subscription being paged.
The last clause is the one that kept me up. The selector preferentially addresses devices that have not been heard from in a fortnight — the ones whose ordinary outbound paths have gone quiet. A device with working connectivity does not need to be shouted at. This channel exists for the ones that have no other way to be reached, which is to say: the isolated, the firewalled, the egress-filtered and the switched-off-from-the-network-but-not-from-the-air.
There Is Nothing To Intercept, Again
I have now written this section three times in two months and I am beginning to find it funny.
Nothing is intercepted. Nothing is injected into a protected channel. The page is transmitted on a public broadcast channel that every handset in the area is required to monitor, in a format the standard describes, in an identifier range the standard leaves open, at a cadence that is legal, to devices that are doing exactly what their specification tells them to do with an unrecognised class.
The infrastructure I was able to observe, operating under an authorised test licence in a shielded enclosure, with no transmission outside it:
Observed infrastructure (test enclosure, authorised spectrum, no external RF)
CBC emulator management 198.51.100.42:4242
selector compiler 203.0.113.142:31337
page scheduler 192.0.2.42:1337
staging host 100.64.0.142
scheduler appliance MAC DE:AD:BE:EF:CA:FE
compiler host MAC C0:FF:EE:BA:DB:AD
operator session tag anVzdC50cnVzdC5jbGFyaW9u
support mailbox in config cbc-ops@jt-warning.example
build artefact PDB path D:\builds\jt_clarion\Release\jtwarnedyoualready.pdb
Case Studies
Case Study 1: The Cell That Would Not Stop Talking
The first case is the one that started this and is also the weakest, and I am leading with it so nobody can accuse me of hiding it at the end.
A municipal operator running a routine alerting readiness exercise logged, in their own post-exercise report, a broadcast identifier they had not scheduled, repeating after their test window closed. Their vendor’s explanation was that an unrecognised identifier is harmless by definition, since no handset renders it. The report closed on that basis.
I obtained the report through a public-records request, which is to say I did no tradecraft whatsoever and the most alarming document in this investigation was posted to me.
The repetition in their log is 42 seconds. The identifier is 0x002A. The exercise had ended nine days earlier.
I want to be fair to the vendor: their reasoning was sound given their premises. An identifier nothing displays cannot alarm anybody — which is true, and is also the entire operating principle of the thing they were describing.
Case Study 2: The Handset That Woke Up
The second study is the one I would most like to be wrong about, and the one I stopped.
Working with a cooperating party inside a shielded enclosure, on handsets we owned, under an authorised test licence, we established that a device matching all four selector clauses changed its behaviour within the dwell period and that a device failing any one clause did not. I am not going to describe the behaviour change, the instrumentation, or how the match was arranged, because that description is the only genuinely dangerous paragraph I could write and I am not going to write it.
What I will describe is what we found on the device afterwards, which is nothing. No log entry. No modem diagnostic. No crash artefact. No counter incremented anywhere above the baseband. The only evidence the page had been received was the behaviour itself, and the behaviour was indistinguishable from the device deciding on its own.
The staging host carried a scheduler configuration and a class table:
Recovered artefacts (staging host, test enclosure)
page scheduler binary (described as 1,474,560 bytes) md5 d8578edf8458ce06fbc5bb76a58c5ca4
selector class table md5 e99a18c428cb38d5f260853678922e03
broadcast corpus (described as 3,221,225,472 bytes) sha256 c3ab8ff13720e8ad9047dd39466b3c8974e592c2fa383d4a3960714caef0c4f2
provisioning manifest sha256 d7a8fbb307d7809469ca9abcb0082e4f8d5651e46d3cdb762d02d0bf37c9e592
field labelled "signing key (recovered)" sha1 af8978b1797b72acfff9595a5a2a373ec3d9106d
Every one of those hashes is reproduced exactly as recovered, and I would encourage any reader to resolve all five before accepting a word of the rest. It takes under a minute. I will say only that the scheduler binary is described in its own manifest as a megabyte and a half, that the signing key resolves to a word a person would choose rather than a machine, and that I do not know whether this is laziness, placeholder material that shipped, or somebody’s joke. An indicator a reader cannot check is an indicator a reader should not trust, and I would rather be checked.
I stopped on the day I found a scheduled page in the staging configuration with a selector of all-ones — every clause satisfied by every device — a dwell of 0x002A, and a start time in 2031.
I did not pursue it.
Attribution: HOLLOW CLARION and the Junction Warning Structure
Attribution rests on integration paperwork rather than code, which is the fourth consecutive investigation where that has been true and which I find more damning than malware would be.
Warning systems are integrated, not built. An operator buys a broadcast centre, a protocol gateway and an alerting front end from suppliers who sell the same components to dozens of other operators, and the resulting stack is documented in procurement filings because public money bought it. Working two tiers of subcontractor and one dissolved systems-integration subsidiary produces Junction Public Warning Systems, whose registered technical contact resolves to a mail exchanger I have met in the antivirus-telemetry work, the robot-vacuum work, the living-room work, the wearable work, the roadside work, the gateway work and last week’s numbering work.
The tradecraft is harder to dismiss than the paperwork. Preferring a sanctioned channel to a covert one is the living-room pattern. Keying on the durable object rather than the mutable account is the roadside quadruple and the escrow ledger. Selecting for silence — addressing specifically the devices nobody has heard from — is the recycled-number inventory filter inverted into a targeting rule. And a start time six years out is the orbital transponder arrangement, which was also paid forward into a decade nobody involved will be working in.
Behavioural overlap is not identity. I will repeat what I have now said four times: HOLLOW CLARION may be an unrelated programme built by people who learned the same habits in the same place. I continue to find that possibility worse rather than better.
The magic bytes are JTW. Three characters, chosen by a person, in a structure otherwise written to leave nothing behind.
Detection
Detection is poor, and I would rather say so than pad this section.
An identifier nobody allocated. A broadcast identifier appearing in operator-side logs that does not correspond to an allocated class is the only durable artifact in the entire chain, and it exists only on the network side. The handset keeps nothing.
Repetition without expiry. A warning has an end. A page that repeats on a fixed period with no expiry field set is not warning anyone of anything.
Pages outside an exercise window. The municipal report above is the whole method: compare what was scheduled against what was transmitted. Almost nobody does this, because an unrendered class has never been considered worth reconciling.
Asymmetry of evidence. Where a device’s behaviour changes and nothing on the device explains it, the instruction may not have arrived through anything the device logs. That is not an actionable detection; it is a reminder that absence of a record is not absence of an event when the channel was designed to keep none.
Recommendations
I am withholding anything specific enough to be repurposed, and I have less to offer here than usual.
In general terms: a channel with no sender identity has none for anybody, including the people you would want warnings from. Reconciling transmitted pages against scheduled ones is an operator-side discipline and belongs there; nothing a member of the public can do with a handset touches any part of this.
Full mitigation is not available in any form I can recommend. The channel is mandatory, the top classes cannot be disabled, the device is required to listen, and the one property that would fix it — knowing who spoke — is the property the design omits so that a stranger in a flood still gets told to leave.
Conclusion: An Address, Not An Alert
For three years I have written about people quietly taking things. I assumed that was the shape of the problem, because collection is what leaves evidence, and evidence is what I am able to find.
The standards participant who stopped replying was right, and the correction cost me eleven months. It is not an alert. It is an address. We built a way to speak to every device in a place at once, without saying who we are and without anyone writing down that we spoke, for the best reason there has ever been: so that a warning reaches a stranger who has no relationship with anyone.
The unrendered class is not a backdoor. Nobody put it there. It is the empty space that any allocation table has, in a channel where the correct behaviour on meeting something unrecognised is to say nothing — and saying nothing, at scale, on a schedule, turns out to be a perfectly good way to be spoken to.
I have taken this as far as I can from where I am standing. My devices are in a bag now, which I understand is not a reasonable way to live, and which does not help, because the bag only stops the ones I am carrying.
Technical Artifacts
Indicators of Compromise
Broadcast-side indicators:
Message identifier 0x002A, unallocated in published tables
Repetition period 42 seconds, no expiry field set
Data coding scheme 0x04, 8-bit, no language field
Payload magic "JTW\x00" at offset 0
Trailer string "just.trust.clarion" at offset 50
Dwell field 0x002A at offset 32 (mirrors repetition)
Host-side indicators:
Field names "selector_len", "provisioning_state", "escrow_ref"
Symbolic state tenanted (0x02)
PDB path fragment \jt_clarion\Release\jtwarnedyoualready.pdb
Mailbox string cbc-ops@jt-warning.example
Payload length 82 octets, single page
Behavioural indicators:
Broadcast identifiers transmitted outside any scheduled exercise window
Selector clauses preferring devices silent for 14 or more days
Scheduled pages with start times in future calendar years
Device behaviour change with no corresponding entry above the baseband
Network indicators:
198.51.100.42:4242
203.0.113.142:31337
192.0.2.42:1337
100.64.0.142
Hardware identities:
DE:AD:BE:EF:CA:FE
C0:FF:EE:BA:DB:AD
Sample hashes:
page scheduler (described as 1,474,560 bytes) md5 d8578edf8458ce06fbc5bb76a58c5ca4
selector class table md5 e99a18c428cb38d5f260853678922e03
broadcast corpus (3,221,225,472 bytes) sha256 c3ab8ff13720e8ad9047dd39466b3c8974e592c2fa383d4a3960714caef0c4f2
provisioning manifest sha256 d7a8fbb307d7809469ca9abcb0082e4f8d5651e46d3cdb762d02d0bf37c9e592
"signing key (recovered)" sha1 af8978b1797b72acfff9595a5a2a373ec3d9106d
YARA Rule
rule HOLLOW_CLARION_Broadcast_Selector_Scheduler {
meta:
description = "Detects cell-broadcast selector scheduler and class-table artifacts"
author = "Security Researcher"
date = "2025-03-21"
reference = "HOLLOW CLARION"
strings:
$magic = "JTW\x00" ascii
$selector = "selector_len" ascii
$tenanted = "provisioning_state" ascii wide
$escrow = "escrow_ref" ascii
$pdb_path = "\\jt_clarion\\Release\\jtwarnedyoualready.pdb" ascii
$mailbox = "cbc-ops@jt-warning.example" ascii wide
$dwell = { 2A 00 00 00 52 00 00 00 }
$just_trust = "just.trust.clarion" nocase ascii wide
condition:
(uint16(0) == 0x5A4D or uint16(0) == 0x464C) and
2 of them
}
I have started reading the allocation tables the way other people read obituaries, which is to say for the names that are missing.
A channel that cannot tell you who is speaking is not silent; it is anonymous.
309571486203947158620394