Skip to content
Emberwire OTA firmware delivery

Releases · 2026

Changelog

Platform releases, agent releases and security advisories. Agent versions are what runs on your devices, so they are the ones worth reading carefully — a platform release you can ignore, an agent release you have to roll out.

Agent support matrix

Agent Released Status Receives Supported until
2.9.x16 Sep 2026 current Features and fixes—
2.8.x9 Jul 2026 supported Fixes and security9 Jul 2027
2.7.x14 May 2026 supported Fixes and security14 May 2027
2.6.x5 Mar 2026 maintenance Security only5 Mar 2027
2.5.x22 Jan 2026 maintenance Security only22 Jan 2027
2.4.x and older2025 and earlier unsupported Nothingended

Deprecations with a date attached

  • manifest_version 1 stops being produced on 1 Dec 2026. Agents below 2.5.0 cannot read version 2; they are also unsupported, which is the same problem stated twice.
  • TLS 1.2 remains accepted for device connections until agent 3.0. There is no date for 3.0 yet, and there will be at least two quarters of notice.
  • The devices?offset= query parameter was removed on 18 Feb 2026. Use next_cursor.

2026

22 September 2026 platform 2026.09.3

Fourth fallback node in candidate lists

Devices whose transfers previously failed on all three offered nodes — almost always units behind carrier-grade address translation with aggressive session timeouts — now receive a fourth candidate. Measured effect on the affected population: transfer completion rose from 96.1 % to 99.2 %.

Also: node health checks now expose cache-hit ratio per product line; wave status returns a per-cohort error-code histogram instead of a total.

16 September 2026 agent 2.9.0

Smaller minimum build, resumable patch application

The minimum bare-metal build dropped from 181 KB to 168 KB by removing the bundled base64 decoder and tightening the manifest parser. Static RAM is unchanged at 24 KB.

Patch application is now resumable: a device that loses power while applying a delta continues from the last completed block instead of re-fetching. On the 1.4 GB terminal line this removes an average of 38 MB of re-transfer per interrupted update.

Also: ew_reject() added as an explicit counterpart to ew_confirm(); jitter is now applied to the first check-in after boot, which matters for fleets that power up together.

9 September 2026 platform 2026.09.1

Nodes remove themselves on bad-block rate

Direct result of the 8 September Warsaw incident. A node that serves blocks failing their digest above a threshold now fails its own health check and drains, instead of waiting for three devices to report E_DIGEST against it.

Incident detail on the status page. No device installed a corrupt image; the change reduces how long devices have to work around a failing node.

27 August 2026 agent 2.8.2

Fixes

  • Backoff timer could overflow on targets with a 32-bit millisecond tick after 49 days of continuous failure. A device in that state stopped retrying entirely.
  • Zephyr build no longer requires CONFIG_HEAP_MEM_POOL_SIZE when crypto is supplied by the platform.
  • E_NOSPACE now reports free bytes on the staging volume rather than on the active slot.
11 August 2026 platform 2026.08.2

Signed audit export

Audit exports are now accompanied by a detached Ed25519 signature over the canonical JSONL, so a regulator can verify the file has not been edited after leaving us. Verification key is published at /v1/exports/key and rotates yearly with overlap.

Requested by metering customers who were previously attaching our exports to filings with no way to demonstrate integrity.

30 July 2026 agent 2.8.1

Fixes

  • A ranged resume across a node change could request one block twice. Harmless, but it showed up as duplicate egress on customer bills; affected months were credited.
  • RISC-V target: misaligned access in the digest routine on cores without the C extension.
  • Linux build: the systemd unit now waits for time synchronisation before the first check-in so log timestamps are usable.
16 July 2026 platform 2026.07.2

Scheduler intent log

After the 14 July hold, the wave scheduler keeps a local intent log and replays it once the database is writable again. A failover now costs the failover time, not the time until someone notices. The scheduler still refuses to advance without a durable record — that behaviour was correct and is unchanged.

9 July 2026 agent 2.8.0

RISC-V, and a smaller I/O path

RV32IMC support for bare metal and Zephyr, with software crypto fallback. The streaming path now hashes directly out of the I/O buffer, which removes one copy and lets 4 KB buffers work on parts where 8 KB previously did not fit.

Also: candidate node list is persisted across reboots, so a device that reboots mid-wave does not have to check in before resuming a transfer.

11 June 2026 agent 2.7.3

Fixes

  • Trial-boot flag was cleared before the metadata journal was flushed on one flash driver, which could mark an unconfirmed slot as permanent after a power cut in a 40 ms window.
  • FreeRTOS: task stack requirement reduced by 2 KB.
20 May 2026 platform 2026.05.2

Webhook worker isolation

Per-destination worker pools and a 10-second response timeout, after the 2 May backlog. One slow endpoint can no longer delay deliveries to anyone else.

14 May 2026 agent 2.7.0

Quarantine cohorts and revocation lists

Devices returning after a long absence are walked through required waypoint releases instead of being offered the newest build directly. Manifests can now carry a revoked_keys list, which the agent persists and applies to every subsequent verification — including offline installs from a technician's laptop.

This is the minimum agent version for the release_counter rollback protection described in Security.

2 April 2026 agent 2.6.1 · EW-2026-002

Security: out-of-bounds read in the patch applier

A manifest containing a malformed delta descriptor could make the patch applier read past the end of its scratch buffer on 32-bit targets, causing a crash and, on some parts, a reset loop until the watchdog reverted the slot.

Not exploitable to install unsigned firmware: the chain and digest checks run before the applier is reached, so producing such a manifest requires a valid release key. Severity assessed as moderate because a compromised signing key already implies worse. Update to 2.6.1 or later; fixed in all supported branches on the same day.

Reported by an independent researcher through security@{{DOMAIN}}, disclosed 21 days after the fix reached 95 % of affected devices.

5 March 2026 agent 2.6.0

Block-level verification while streaming

The agent now verifies each block as it arrives instead of only hashing the staged slot at the end. A corrupt block is re-fetched immediately, which turns a full re-download into a 64 KB one. This is the change that made the September Warsaw incident a slowdown rather than a fleet-wide re-transfer.

18 February 2026 platform 2026.02.3

Cursor required on device search

Unbounded device searches now return E_SELECTOR. Follows the 11 February latency incident. The offset parameter was removed in the same release; it had been deprecated since the previous quarter.

12 February 2026 agent 2.5.4 · EW-2026-001

Security: manifest parser accepted a duplicate field

A manifest containing the digest field twice was parsed last-wins while the signature covered the canonical form, creating a theoretical path to staging an image whose digest was not the signed one. No such manifest was ever observed, and producing one still required a valid release key.

Fixed by rejecting any duplicate key in a manifest object outright. Found during our own fuzzing of the parser.

22 January 2026 agent 2.5.0

Bandwidth ceiling and device-local maintenance windows

A per-device transfer ceiling, and window evaluation in the device's own timezone rather than in the operator's. Both came from router customers whose users noticed the updater competing with their traffic in the evening.

15 January 2026 platform 2026.01.1

Twelve delta bases kept hot

Raised from six. Fleets with slow-moving devices were falling off the delta path and pulling full images; at twelve bases the share of devices receiving a full image on a typical release dropped from 9.4 % to 4.1 %.

Across the customer base this is roughly 70 TB a month of egress that is no longer sent — most of it to devices on metered connections.

How to read this

Platform releases

Applied by us, no action needed. Behavioural changes are listed; if one could change what your devices receive, it says so explicitly and is announced on Status beforehand.

Agent releases

You roll them out like any other firmware change — through Emberwire, in waves, with the same halt rules. The agent is part of your image, not something we can update behind your back.

Advisories

Numbered EW-YYYY-NNN, published once a fix has reached most affected devices, with the reasoning for the severity we assigned. Two in 2026, both requiring a valid signing key to reach.