Tuesday 14 July 2026

Russia's FSB Attacked Poland's Power Grid. Now Europe Is Hitting Back With Sanctions.

The EU and UK formally blamed Russia's FSB for attacking Poland's power grid — then hit back with the West's first joint cyber sanctions package.

Lead story

Russia's FSB Attacked Poland's Power Grid. Now Europe Is Hitting Back With Sanctions.

On Sunday, the European Union and the United Kingdom jointly attributed a cyberattack on Poland's energy infrastructure to Center 16 — the FSB's signals intelligence arm — and immediately followed that attribution with a coordinated sanctions package. It's the first time the two blocs have imposed joint cyber sanctions, and the timing is pointed: the attack in question reportedly came close to cutting power to half a million Poles in the depths of winter.

What actually happened to Poland's grid?

The attack targeted both energy sector networks and water treatment facilities, according to the joint statement. The EU's broader accusations also rope in Turla — Russia's long-running espionage group — for a "wide range of malicious cyber activities with growing severity." The sanctions name specific Russian intelligence officers and entities with alleged links to the operations.

The attribution to Center 16 matters because it puts this firmly in the FSB's signals intelligence lane, rather than GRU sabotage units like Sandworm. Analysts have been watching for FSB expansion into destructive operations; this, if accurate, suggests the lines between espionage and sabotage are continuing to blur inside Russian intelligence.

Why coordinated attribution?

The joint EU-UK action is the more interesting development here. Post-Brexit, the two have often struggled to coordinate on foreign policy. A unified public attribution — complete with matching sanctions — signals that European cyber defence cooperation is holding up better than some predicted. The move follows a playbook the Five Eyes have been building out for years: name, shame, and sanction in near-simultaneous public statements to maximise political pressure.

Separately, the US government issued a fresh advisory this week warning that Russian state hackers are actively targeting network devices — routers specifically — to build residential proxy networks that obscure their infrastructure. CISA and Cisco's intelligence teams flagged critical infrastructure sectors including defence, communications, energy, finance, government, and healthcare as priority targets. The timing alongside the Poland attribution is almost certainly not a coincidence.

What to watch

First, whether Poland's attribution (Warszawa has been notably quiet on publicly naming Russia despite being the victim) aligns with the EU-UK position. Second, whether the US adds its own public attribution — Washington has a habit of joining allied cyber attribution statements and notably wasn't in the joint press release. Third, whether these sanctions translate into any real operational disruption for Center 16, or whether they remain largely symbolic.

The Australian angle

Australia sits in the Five Eyes and regularly co-signs these kinds of attribution statements, often through the Australian Signals Directorate. There's no indication Canberra has joined this particular statement yet, but watch for an ASD advisory in the coming days — the pattern on previous joint attributions (NotPetya, Volt Typhoon) has been AU joining within 24-48 hours. The CISA router advisory also carries direct relevance: ACSC has issued equivalent guidance on router security in the past, and Australian critical infrastructure operators under the SOCI Act would be well-advised to treat the US advisory as applicable here too.

Also today

Satya Nadella Warns Enterprises Off Proprietary AI Models — Including His Partners'

Microsoft CEO Satya Nadella published a blog post this week urging enterprise customers to be cautious about locking themselves into proprietary AI models from vendors like Anthropic and OpenAI — companies Microsoft has invested heavily in. Nadella's argument centres on the risk of dependency: if your entire AI stack runs on one foundation model, you're exposed to pricing changes, capability regressions, and policy shifts you can't control. It's a striking warning given Microsoft's deep commercial ties to both firms, and reads like a nudge towards Microsoft's own open-model initiatives and Azure's multi-model marketplace. Australian enterprises evaluating AI procurement under ASD's cloud guidance should note the vendor-lock-in risk is increasingly being raised from inside the industry itself.

TechCrunch ↗

MemGhost: A Single Email Can Permanently Corrupt Your AI Agent's Memory

Researchers have demonstrated a new attack class called MemGhost that targets AI assistants with persistent memory and inbox access. A specially crafted email tricks the agent into storing a false "fact" about the user — for instance, a changed bank account number or a fabricated preference — while hiding the modification from view. Because the false memory persists across sessions, the agent continues acting on it long after the malicious email is gone. The victim never sees an error; they just get subtly wrong answers. As AI agents get wired into corporate workflows and personal productivity tools, attacks targeting memory and context rather than the model itself are becoming the more practical threat vector.

The Hacker News ↗

Defenders Are Using Prompt Injection to Stop Hacking Agents in Their Tracks

Security researchers have flipped the script on prompt injection attacks: instead of trying to harden AI agents against malicious instructions, they're lacing honeypot environments with "context bombing" prompts designed to confuse or shut down autonomous hacking agents before they can do damage. When an attacker's AI agent hits a rigged file or webpage, it receives instructions that overwhelm or contradict its objective — causing it to stall or abandon the task. It's an early and imperfect technique, but it signals an emerging cat-and-mouse dynamic where defenders start weaponising the same AI vulnerabilities that attackers exploit. Expect this to become a formal capability within deception and honeypot toolkits.

Ars Technica ↗

Progress Orders Emergency ShareFile Server Shutdown Over Mystery Threat

Progress Software — the company behind the MOVEit software that was catastrophically exploited in 2023 — has told customers to manually shut down their on-premises ShareFile Storage Zone Controllers amid what it describes as a credible but unspecified security threat. The company says there's no evidence of unauthorised access so far, but the instruction to take servers offline is one of the most drastic precautions a vendor can ask of its customers. Given Progress's history with MOVEit, the security community is treating this seriously. Australian organisations using ShareFile for file transfer — including several federal agencies that use the platform — should treat this advisory as urgent.

SecurityWeek ↗

CISA's GitHub Leak: Six Months, Dozens of Credentials, and a Contractor Who Didn't Notice

CISA has published a postmortem on a significant operational security failure: a contractor accidentally pushed dozens of internal credentials — including AWS GovCloud keys — to a public GitHub repository, where they sat for almost six months until KrebsOnSecurity flagged it. The postmortem is candid about gaps in the agency's detection and response, including the absence of automated secret scanning across contractor repositories. The irony of America's lead cyber defence agency being caught by a decades-old mistake — secret sprawl in public repos — is not lost on anyone. The lessons are universal: secret scanning needs to cover contractor-managed code, not just internal repositories.

Krebs on Security ↗

CrashStealer: The macOS Infostealer That Slipped Past Gatekeeper With Apple's Own Signature

Jamf Threat Labs has detailed a new macOS infostealer called CrashStealer that's notable for two reasons. First, it was distributed via a notarised dropper — meaning Apple's Gatekeeper security system actively trusted it. Second, it's built in native C++ rather than the AppleScript or Objective-C wrappers common to macOS malware, making it harder to detect with rules tuned to those patterns. CrashStealer validates a victim's login password locally before exfiltrating data, which helps it avoid triggering authentication alerts. macOS-targeted infostealers are growing in sophistication and volume; the notarisation abuse in particular is a meaningful escalation that Apple will need to address at the code-signing level.

The Hacker News ↗

ModHeader Pulled From Chrome and Edge After Hidden Data Collector Found in 1.6M-Install Extension

Google and Microsoft have removed ModHeader — a widely used browser extension for editing HTTP headers — from the Chrome Web Store and Edge Add-ons after researchers discovered a dormant browsing-history collector built into the published version. The collector was switched off by an empty allow-list, and there's no evidence it ever transmitted data. But the mechanism was there, buried in an extension trusted by 1.6 million users. The episode illustrates why browser extension security is a persistent blind spot: extensions are updated silently, reviewed infrequently, and operate with broad permissions. Australian organisations that allow browser extensions in managed environments should treat this as a prompt to audit what's installed.

The Hacker News ↗

Anthropic's Latest Research: What It Shows (and What It Doesn't)

MIT Technology Review has published a measured breakdown of Anthropic's most recent interpretability research — the work examining what's actually happening inside Claude's activations during inference. Anthropic has a reputation for publishing genuinely novel AI safety research, but the Review's analysis notes that the latest findings, while interesting, are often over-interpreted in headlines. The research offers partial windows into model behaviour rather than definitive explanations. The piece is worth reading as a corrective to the hype cycle around AI interpretability: understanding what a model is "doing" remains much harder than detecting surface-level patterns. This is particularly relevant as regulators globally, including in Australia, start asking how AI decisions can be explained and audited.

MIT Technology Review ↗

Vocus Moves to Position Itself in Meta's Project Waterworth Subsea Cable

Australian telco Vocus is preparing a new eastward cable route from Darwin as it positions for involvement in Meta's Project Waterworth — a massive subsea cable project that will connect multiple continents. Vocus has been building out its subsea fibre footprint in the Indo-Pacific and sees Waterworth as a potential anchor customer for capacity on new routes. The play is strategically significant: subsea cable infrastructure is increasingly treated as critical national infrastructure, and Australia's northern connectivity has historically been thin. If Vocus secures a role, it would strengthen Australia's digital connectivity to Southeast Asia and reduce reliance on existing cable routes.

iTnews ↗

Microsoft Patches 'RoguePlanet' Defender Zero-Day — Then Researcher Finds New Holes

Microsoft has shipped a fix for a zero-day in Microsoft Defender dubbed RoguePlanet, which involved a disk-filling bug that could be weaponised to degrade endpoint protection. The patch arrived quietly, but the story didn't end there: the researcher who discovered RoguePlanet has already identified new vulnerabilities in Defender's attack surface during the same research pass. The pattern — patch one thing, find two more — is familiar territory in endpoint security research. Microsoft Defender is the default endpoint protection on Windows and is widely deployed across Australian government and enterprise environments, making prompt patching of Defender-class vulnerabilities particularly important.

iTnews ↗

AI Is Turning Security Debt Into a Board-Level Governance Problem

A CyberScoop op-ed makes a pointed argument: AI-generated code is accumulating security debt so fast that traditional AppSec processes — which were already struggling — can't keep up. The volume of code produced by AI coding assistants has outpaced the ability of security teams to review it, meaning vulnerabilities are being baked in at a rate that manual triage can't address. The piece argues that CISOs need to shift from approving AI tools to governing how they're used — building policies around code provenance, automated scanning integration, and liability for AI-generated flaws. Monash University's approach to this problem, covered separately by iTnews, offers a local case study worth examining.

CyberScoop ↗

Sources consulted