Daily brief at 7am Melbourne. Unsubscribe any time.

Thursday 20 August 2026

The Ghost in Cloudflare's Machine: Spectre Is Back, and It's Faster Than Ever

A Spectre-variant attack is leaking secrets from Cloudflare's production edge network — and it's 360 times faster than the last one.

Lead story

The Ghost in Cloudflare's Machine: Spectre Is Back, and It's Faster Than Ever

Researchers have demonstrated a working Spectre attack against Cloudflare Workers — the serverless compute platform running at the edge of the internet — that successfully extracted a JSON Web Token from a co-located Worker in a live production environment. The kicker: it does so at up to 12 bits per second, which sounds glacial until you realise it's 360 times faster than the previous benchmark set in 2021.

Spectre, for the uninitiated, is a class of hardware vulnerability disclosed back in 2018 that exploits the way modern processors speculatively execute code. The short version: a malicious process can trick a CPU into briefly touching memory it shouldn't, then measure the timing of that access to infer what was there. Chip makers and cloud vendors spent years patching, mitigating, and generally hoping the problem would quietly go away. It hasn't.

What makes this finding significant is that it works against a hardened, real-world production environment — not a lab simulation. Cloudflare Workers isolates tenants using V8 isolates rather than traditional virtual machines, which is faster and cheaper but places more trust in software-level sandboxing. The researchers used an attacker-controlled Worker and a victim Worker side by side, the same way two unrelated customers might be co-located on Cloudflare's infrastructure, and leaked a JWT — a credential used to authenticate API calls — from the victim's memory.

JWTs are valuable targets. They're the keys that say "this request is authorised," and stealing one lets an attacker impersonate the legitimate service until the token expires or is revoked. In a real attack scenario, the leaked token could be used to make authenticated API calls, pivot deeper into a target's infrastructure, or access data the attacker was never meant to see.

Cloudflare has been notified and is working on mitigations. The company has long acknowledged that shared-hardware isolation is a hard problem, and has previously invested in techniques like timer resolution reduction and process isolation to blunt Spectre-style attacks. But the research suggests those defences haven't kept pace with attacker ingenuity.

The broader implication is uncomfortable for the entire serverless and edge-compute industry. The economics of shared infrastructure — many tenants, one machine — are fundamentally in tension with the assumption that tenant workloads are fully isolated. Every major cloud provider runs some variant of this model. AWS Lambda, Google Cloud Run, Azure Functions: all of them are betting that software isolation is good enough. This research is a data point suggesting it may not be.

For Australian organisations using Cloudflare Workers — and there are many, given Cloudflare's widespread deployment across Australian enterprise and government — the immediate action item is to rotate JWTs and other short-lived credentials regularly, treat edge compute as a partially-trusted environment, and avoid storing high-value secrets in Worker memory any longer than necessary.

The full research paper is expected to be published shortly. Watch for Cloudflare's formal response, which will likely include changes to timer resolution and potentially new guidance on sensitive workloads.

Also today

China-Linked Operator Runs 'Near-Autonomous' AI Attack on APAC Government

A Chinese-language threat actor has reportedly conducted what researchers are calling the first near-autonomous AI-assisted attack on a nation-state target, likely Taiwan. The operator used a complex AI framework to identify, target, and compromise government agencies with minimal human intervention at each stage. If confirmed, it marks a qualitative shift in how state-linked actors are using AI — moving from AI-assisted reconnaissance to AI-directed execution. Australian government agencies operating in the Indo-Pacific context should treat this as a live signal of where nation-state capability is heading, particularly given Australia's own regional exposure.

Dark Reading

NSA and FBI Warn AI-Generated Tools Are Hitting Critical Infrastructure PLCs

The NSA, FBI, and partner agencies have issued a joint warning about a campaign targeting Siemens S7 Series programmable logic controllers — the hardware that runs water treatment plants, power stations, and industrial facilities. The attackers are using AI-assisted tooling to accelerate exploit development alongside known vulnerabilities. This appears to be the first time US agencies have explicitly cited AI-generated attack tools in an active critical infrastructure campaign. Australia's water and energy sectors use the same Siemens PLC families, and the ASD's ACSC has previously flagged critical infrastructure as a priority target for state-linked actors.

The Record

Quest Hotel Chain Guests' Data Exposed After Third-Party Breach

An Australian hotel chain — Quest, with over 120 properties across the country — has had guests' personal information exposed after a breach at a third-party database operator. The incident is a textbook third-party risk scenario: the chain itself wasn't compromised directly, but a vendor holding its customer data was. Under Australia's Privacy Act and the Notifiable Data Breaches scheme, the obligations flow to the entity that collected the data, not just the one that lost it. It's also a timely reminder that the SOCI Act's third-party risk provisions exist precisely for situations like this.

The Register

Cl0p Names 40+ Victims in PTC Windchill Campaign — Including Shell and Philips

The Cl0p ransomware group has published a list of more than 40 organisations it claims to have compromised via a zero-day in PTC Windchill and FlexPLM — product lifecycle management software used heavily in manufacturing and engineering. Named victims include Shell, Philips, Fiserv, and Zebra Technologies. Separately, ReliaQuest found the JSP web shell deployed in these attacks is purpose-built for Windchill environments, capable of mapping engineering vaults and decrypting credentials. Cl0p reportedly sat inside victim networks for roughly a month before sending extortion emails — the long dwell time is the story within the story.

SecurityWeek

SilkParasite: A New Chinese-Nexus Espionage Cluster Targets Central Asian Governments

A previously undocumented espionage operation called SilkParasite has been caught targeting government bodies across Central Asia, deploying seven remote access trojan families — five of which are brand new to researchers. The group is assessed to have a Chinese nexus and links to the FamousSparrow cluster, suggesting an established threat actor has stood up a new operational toolkit to avoid detection. Spear-phishing is the initial access vector. The disclosure underscores how prolific Chinese state-linked APTs have become at fielding novel malware families to stay ahead of signature-based defences.

The Hacker News

CareCloud Health Breach Is Ten Times Bigger Than First Reported — 3.7 Million Affected

Electronic health record company CareCloud has revised the scope of a previously disclosed breach from roughly 350,000 affected individuals to 3,756,469 — a tenfold increase that places it among the larger US healthcare data incidents on record. A hacker spent eight hours inside one of CareCloud's EHR environments before being detected. The delayed and dramatically revised disclosure is a familiar pattern in healthcare breaches, and will likely draw scrutiny from US regulators. Australian healthcare providers using US-based software vendors should note that equivalent obligations under the Privacy Act and NDB scheme apply regardless of where the vendor is headquartered.

The Record

Critical GitLab Zero-Click Flaw Leaves Self-Managed Deployments in the Dark

A critical zero-click vulnerability in GitLab (CVE-2026-19478) has been disclosed with limited technical detail, creating a difficult position for organisations running self-managed GitLab instances. Without enough information to write detection rules or hunt for indicators of compromise, defenders are largely flying blind as to whether they've already been hit. GitLab is widely deployed in Australian enterprise and government DevOps pipelines, including within defence and critical infrastructure environments. The advice for now: apply the patch immediately and treat any anomalous CI/CD pipeline activity in the lookback window as suspicious.

Dark Reading

'Kriminal' AI Platform Offers Guardrail-Free Hacking Tools for Crypto

A new AI platform called Kriminal is offering no-filter access to social engineering assistance, offensive cybercrime tools, and OSINT scanning — all available for cryptocurrency. The company's terms of service nominally prohibit illegal use, a fig leaf that fools nobody. What's notable here isn't that jailbroken AI models exist (they do), but that someone has productised the offering with a polished platform, lowering the barrier for less-skilled threat actors. It's the cybercrime-as-a-service model applied directly to AI — and it's another data point in the accelerating commoditisation of AI-assisted attacks.

Dark Reading

Flock Safety's Next-Gen AI Goes Far Beyond Licence Plates

WIRED obtained and reconstructed code from Flock Safety's new AI surveillance system — already deployed by some US police departments — and found it goes substantially further than the licence plate readers the company is known for. The system can correlate vehicle movement, appearance, and behaviour patterns to build detailed profiles. The disclosure comes as Flock's CEO apologised for documented police misuse of the existing platform, and as activists are reportedly planning Halloween-themed vandalism campaigns against the cameras. The episode is a useful case study in how surveillance infrastructure expands incrementally until it's something quite different from what was originally sold.

WIRED Security

AI Approval Ratings Are Falling Even as Adoption Climbs

A TechCrunch analysis finds that despite AI tools becoming harder to avoid in everyday life, consumer trust in the technology is declining rather than growing. Silicon Valley bet that getting people to use AI would naturally lead to acceptance — the 'try it and you'll love it' theory. The data increasingly suggests otherwise: familiarity is breeding wariness, not enthusiasm. For the industry, this creates a genuine strategic problem. Adoption metrics look good in board decks; sentiment metrics do not. The gap between usage and trust is widening, and that's the number that matters long-term.

TechCrunch AI

Operation CameraSwarm Compromised 14,500 Dahua Security Cameras in Five Weeks

Researchers at Hunt.io have reconstructed a campaign that compromised more than 14,530 Dahua IP cameras between mid-June and late July 2026, using credential stuffing, two authentication-bypass flaws, and a peer-to-peer relay technique to route traffic through the compromised devices. The attackers left behind an exposed working directory containing over 2,600 files — which is how researchers pieced together the operation. Dahua cameras are widely deployed in Australian commercial and residential settings. If you're running Dahua hardware on an unpatched firmware version, treat it as compromised until proven otherwise and check for unexpected outbound P2P connections.

The Hacker News

Sources consulted