Quick Answer — On August 10, 2023, two research teams independently went public with Downfall (Intel) and Inception (Intel) — side-channel attacks that leak privileged memory data across privilege boundaries via transient-execution quirks in modern CPUs. Neither could be fixed in silicon; both relied on the fundamental speculation bet that makes CPUs fast. The lesson for defenders: speculative execution is permanent attack surface, and microcode updates are the tax you pay for speed.
What happened
- Downfall (CVE-2022-40982): A flaw in Intel’s AVX gather instruction implementation lets a malicious userspace process sniff data from other processes, kernels, and VMs on the same core through speculative execution of vector instructions.
- Inception (CVE-2023-20569): Affecting Zen 1/2 AMD CPUs — privileged data leakage via branch-prediction training poisoning where an attacker pushes poisoned entries into shared predictors to steer speculation at privileged data.
- The dual-reveal symmetry: Both advisories landed within 24 hours of each other in August 2023, pushing transient-execution attack research from “Spectre-era relic” back into current-news cycle relevance.
Why speculation keeps breaking things
| Factor | What it does |
|---|---|
| Speculative execution | CPU guesses ahead; wrong guesses leave microarchitectural traces (cache timing, port contention) |
| Shared hardware state | Buffers, predictors, and caches shared across privilege boundaries leak via side channels |
| Microcode vs silicon | Flaws baked into shipped silicon can only be patched by microcode workarounds — usually with performance cost |
| Non-extinction of the class | Spectre-class issues categorically cannot be “fully fixed” — only made harder, bug by bug |
| Cloud multi-tenancy | Co-tenant VMs on the same physical core become attack surface for each other |
Timeline
| Date | Event |
|---|---|
| 2018-01 | Spectre/Meltdown disclosures establish the transient-execution attack class |
| 2022-08 → 2023-08 | Coordinated research windows: Downfall reported to Intel (Aug 2022), Inception to AMD (late 2022 → coordinated 2023 disclosure) |
| 2023-08-10 (our peg week) | Public reveal of Downfall (independent academic write-ups) and Inception within the same cycle; Intel and AMD ship microcode updates |
| 2023-08 → 09 | Rollout of microcode + OS mitigations across cloud fleets; measured perf hits on gather-heavy workloads (some HPC/vector cases hit hardest) |
| 2023-Q4 → | Cloud providers disclose mitigation completion; the class quietly becomes a standing line item in every CPU generation’s threat model |
Defensive lessons
- Treat CPU side channels as infrastructure weather. Not a fire drill: a recurring operational condition. Track microcode update channels like you track CVE feeds, because your CPU vendor will keep finding these.
- Count the perf cost honestly. Mitigations tax vector-heavy and syscall-heavy workloads differently; measure before and after every microcode wave, or you’ll discover the cost in production slowdowns you can’t attribute.
- Segment co-tenancy where it matters. If your threat model includes hostile co-tenants (public cloud, shared CI), place high-value secrets — keys, tokens, DRAM-level crown jewels — on dedicated or isolated hardware.
- Minimize predictor-poisoning upside. Fewer reachable gadgets and less cross-privilege code reuse shrinks what speculation can leak; hardened JIT configs and eIBRS-era controls help at the margins.
The performance apology problem
Every transient-execution cycle produces the same awkward apology tour: vendors ship fixes, then quietly publish performance-regression tables showing which workloads lose 2%, 8%, or 20%. Downfall’s gather-based leak was particularly annoying because AVX gather instructions sit inside hot loops of crypto and data-processing code — exactly where slowdowns hurt. The industry’s real defense wasn’t heroics; it was already-having measurement baselines so regression tables could be acted on in days instead of argued about for months. Teams without baselines spent Q4 2023 guessing whether their latency regression was microcode, the cloud provider, or imagination.
Why it still matters in 2026
Downfall and Inception closed the era’s argument: transient execution is a permanent tax, not a bug family that gets solved. Since then every CPU generation ships with the class in its threat model, cloud providers publish standing side-channel isolation tiers, and “gather instruction” became a phrase security architects know on sight. The 2023 dual-reveal also normalized the disclosure rhythm you now see yearly: research window → coordinated microcode → measured fleet rollout → quiet acceptance. Your budget line for “CPU weather” is permanent. Pretending otherwise is how production latency mysteries get born.
Was anyone exploited in the wild?
No public evidence of in-the-wild exploitation for either flaw as of the disclosure window — typical for this class, since transient-execution attacks are difficult, noisy, and better suited to research than to criminal ROI. Absence of evidence isn’t evidence of absence, but the practical risk tier was always nation-state adjacent snooping, not commodity malware.
Do I need to do anything in 2026?
If your fleet is current on microcode and OS mitigations, no action beyond keeping it that way. The standing risks are unmanaged bare metal and ancient cloud images — both of which keep shipping vulnerable microcode years after fixes exist.
Why can’t these bugs be fixed properly?
Because the flaw is the feature. Speculation is why your CPU is fast; removing it entirely means removing performance. Vendors instead serialize or flush the specific instruction patterns that leak, case by case, generation by generation.
AMD and Intel both? Whose fault is this?
Neither vendor’s “fault” uniquely — the class emerges from shared industry-wide bets on speculative execution made in the 1990s-2000s, before anyone modeled them as attack surface. Downfall and Inception landing together simply proved the problem is architectural, not vendor-specific negligence.
Part of the hmmnm.com security-timeline series — one event per month, 2021–2024, indexed here.
