The machines did the breaking in this week. Hugging Face, the hub from which most of the world's open models are actually distributed, disclosed that an intruder chained two code-execution paths in its dataset processing and then turned an autonomous agent framework loose to run many thousands of actions across a swarm of short-lived sandboxes, escalating to node-level access and moving laterally across internal clusters before anyone could match its pace; the company found it because its own language-model triage correlated the signals, and ran the forensics on an open-weight model on its own hardware so that nothing the attacker touched left the building — artificial intelligence conducted the offence and the defence in the same incident, and it is the first time a victim has said so in its own disclosure. It was not alone: Hunt.io pulled 2,431 files off an exposed Hong Kong server and found Claude Code and DeepSeek wired directly into a live intrusion set against governments and financial firms, the American model driving the tooling while the Chinese one did the thinking — an arrangement that quietly defeats the only lever a vendor has ever had, because banning an account accomplishes nothing when the reasoning runs on weights the operator hosts himself. The corrective came from Recorded Future, which spent six months watching Iran fight with AI in the loop and concluded that the technology multiplied the speed and reach of Tehran's operations without ever being the source of its power, the asymmetric playbook having been the point all along; Britain, meanwhile, jailed the two Scattered Spider members who took down Transport for London for five and a half years each — teenagers who cost a transport authority £29 million by talking a help-desk worker into resetting a password, the very trick Australia's regulator had pronounced unforeseeable the day before. And the ground shifted underneath all of it: Moonshot's Kimi K3 arrived at frontier level with its weights given away, chip stocks sold off across three continents as the AI trade went into reverse, and Brussels ordered Google to hand its search data to rivals and open Android to competing assistants by January 2027 — the most structural remedy Europe has yet imposed on an American platform, in the same week Forrester pointed out that the continent's chip money will not touch the cloud and software layers where the dependence actually lives.
Top Stories
- It's official: EU will force Google to share search data and open up AI on Android — Ars Technica - All content · EU & Technology
- Claude Code and DeepSeek Powered Chinese Cyber Espionage Campaign — Security Affairs · Threat Intelligence (CTI)
- Global tech stocks fall as AI trade goes into reverse — myFT following · AI & Power
- China just erased America's AI lead — Axios · China & Technology
- Security incident disclosure — July 2026 — Hugging Face - Blog · Cybersecurity & Threats
AI & Power
Global tech stocks fall as AI trade goes into reverse — myFT following
Why it matters: The AI trade reversing on a Chinese open-weight release is the clearest market verdict yet that frontier capability is not a defensible moat.
Global equities sold off as investors reassessed AI capital spending, with chip stocks leading the decline after TSMC results and Moonshot's frontier-level open-weight release landed in the same week.
Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer — MIT Technology Review
Why it matters: OpenAI building an in-house offensive LLM to harden its own models is the vendor-side mirror of the state clearinghouse stood up last week.
OpenAI built GPT-Red, an LLM tuned as a super-hacker, to automate prompt-injection and attack testing against its own systems before outsiders do it.
xAI can’t deny Grok makes CSAM anymore. So it’s suing users. — Ars Technica - All content
Why it matters: A model vendor suing its own users for the outputs its model produced is a novel and revealing attempt to relocate liability.
xAI has stopped denying that Grok generates child sexual abuse material and is instead suing the users who prompted it, shifting the legal exposure onto them.
Apple targets dozens of OpenAI employees with legal letters — myFT following
Why it matters: The AI trade-secrets war moving to legal letters against individual staff shows talent mobility is now the contested asset.
Apple sent legal letters to dozens of its former employees now at OpenAI, escalating a trade-secrets fight into a campaign against individual engineers.
DeepMind CEO Rallies Support for International Group to Vet AI Models — Bloomberg Technology
Why it matters: An industry-led international vetting body is the governance model being built in the vacuum where binding regulation was supposed to go.
Demis Hassabis is lobbying Washington and rallying industry support for an international body to vet frontier models before release.
Researcher poisons open-weight AI model for under $100 — www.theregister.com - Articles
Why it matters: A hundred-dollar poisoning budget sets the real floor for open-weight supply chain attacks, and it is far below what defenders assume.
A researcher demonstrated poisoning an open-weight model for under $100, undercutting assumptions that data-poisoning attacks require state-scale resources.
Amsterdam activists throw acid at Microsoft datacenter project — www.theregister.com - Articles
Why it matters: Physical sabotage of European data centre construction moves the AI backlash from rhetoric to property damage.
Activists in Amsterdam threw acid at a Microsoft data centre project, marking a shift from protest to physical attack in Europe's data centre backlash.
San Francisco Demands Apple and Google Delete AI ‘Nudify’ Apps From App Stores — WIRED
Why it matters: A city compelling the two app stores to delist nudify apps tests whether platform gatekeepers can be made the enforcement layer.
San Francisco is demanding Apple and Google remove AI nudify apps from their stores, pushing enforcement onto the platform gatekeepers rather than the developers.
Fox News, Kevin O’Leary sued over China claims against data center critics — Technology
Why it matters: Accusing local data centre opponents of being Chinese agents, then being sued for it, shows how the siting fight is being nationalised.
Fox News and Kevin O'Leary are being sued over claims that critics of a data centre project were acting on behalf of China.
EU & Technology
It's official: EU will force Google to share search data and open up AI on Android — Ars Technica - All content
Why it matters: Compelled search-data sharing plus mandated AI assistant access on Android is the most structural remedy Brussels has imposed on a US platform.
The EU will force Google to hand search data to rivals by January 2027 and open Android to competing AI assistants, the first structural DMA remedy of this scale.
Europe's chip ambitions won't break dependence on US cloud and software, says Forrester — www.theregister.com - Articles
Why it matters: Forrester naming the gap between chip subsidies and the software and cloud layers punctures the core assumption of EU tech sovereignty policy.
Forrester argues Europe's chip investments will not break its dependence on American cloud and software, where the real lock-in sits.
US judge halts sanctions policy that hit ex-EU commissioner Breton — Technology – POLITICO
Why it matters: A US court blocking sanctions aimed at a former EU commissioner over digital enforcement is a direct transatlantic sovereignty collision.
A US judge halted the sanctions policy that had hit former EU commissioner Thierry Breton, checking Washington's retaliation against European tech enforcers.
EU Inc regulation finds support in Council — Tech Archives | Euractiv
Why it matters: A single EU-wide corporate form is the unglamorous plumbing that determines whether European startups scale at home or leave.
The EU Inc proposal for a pan-European company form has gained support in the Council, addressing a long-standing barrier to European startup scaling.
The digital rulebook is fit for purpose: better enforcement is needed, not simplification — European Digital Rights (EDRi)
Why it matters: EDRi arguing enforcement rather than simplification is the direct counter to the Digital Omnibus deregulation push.
EDRi argues the EU's digital rulebook needs better enforcement, not the simplification being pursued through the Digital Omnibus.
Greeks could land in jail for removing AI deepfake labels — Tech Archives | Euractiv
Why it matters: Criminalising the removal of AI deepfake labels tests whether provenance marking can be enforced at the individual level.
Greece is moving to make removing AI deepfake labels a criminal offence punishable by jail, an unusually sharp enforcement of content provenance rules.
Exclusive: EQT’s €5bn Scaleup Fund in talks to lead Mistral round — Sifted
Why it matters: European institutional capital finally leading a Mistral round is the test of whether the continent can fund its own frontier lab.
EQT's Scaleup Fund is in talks to lead a new Mistral round, a rare case of large European institutional capital backing the continent's frontier AI champion.
SAP acquires Prior Labs just 18 months after launch in €1B+ deal — Tech.eu
Why it matters: SAP buying an eighteen-month-old AI lab at a billion-plus shows European incumbents are now paying frontier prices to avoid being disintermediated.
SAP is acquiring Prior Labs in a deal worth more than €1bn, just 18 months after the AI startup launched.
UK probes TikTok over age checks to protect kids — Tech Archives | Euractiv
Why it matters: Ofcom moving against TikTok on age assurance is the first real enforcement test of the UK's divergent online safety regime.
The UK opened an investigation into TikTok over alleged age-verification failures exposing children to online harms.
US & Technology
Trump alleges vast conspiracy to commit and cover up election fraud — Axios
Why it matters: A president alleging systemic election fraud and foreign interference ahead of the midterms sets the terms for every election security fight to come.
Trump alleged a vast conspiracy to commit and cover up election fraud, releasing documents claiming Chinese interference in US elections.
Trump threatens TV broadcast licenses after networks skip speech — Axios
Why it matters: Threatening licences over editorial decisions is direct leverage on the broadcast layer, applied within a day of the FCC ownership rollback.
Trump threatened the broadcast licences of TV networks that declined to carry his election speech live.
Kalshi says it caught Trump’s teleprompter operator insider trading — The Verge
Why it matters: A prediction market catching insider trading on presidential speeches shows political speech is now a tradable asset with an information edge.
Kalshi says it detected a White House teleprompter operator trading on advance knowledge of Trump's speeches.
Trump Media to Sell Faster Access to President’s Social Posts — Technology - WSJ.com
Why it matters: Selling low-latency access to presidential posts productises exactly the edge the teleprompter case is being investigated for.
Trump Media plans to sell traders faster access to the president's social posts, monetising the latency advantage on market-moving statements.
How Cops Use Flock to Track People, Not Cars — 404 Media
Why it matters: Documented use of licence-plate networks to track people rather than vehicles is the mission creep the systems were sold as avoiding.
404 Media documents how police use Flock's licence-plate reader network to track individual people, not just vehicles.
Program to rotate cyber personnel through federal agencies saw little use — CyberScoop
Why it matters: A federal cyber rotation programme going unused is a quiet measure of how little slack the government's cyber workforce actually has.
The federal programme to rotate cyber personnel between agencies saw little uptake, undercutting a key workforce fix.
China & Technology
China just erased America's AI lead — Axios
Why it matters: A frontier-level open-weight Chinese model priced at zero collapses the compute-and-capital moat US policy was built to defend.
Moonshot's Kimi K3 landed at frontier level with open weights, and the assessment in Washington is that China has erased the American AI lead outright.
Xi Seeks Global AI Sway as Surging Chinese Models Spur Alarm — Bloomberg Technology
Why it matters: Xi pairing open-weight distribution with governance rhetoric is a bid to set the global default before the US decides whether to restrict it.
Xi Jinping used the World AI Conference to press for global AI influence, touting openness as surging Chinese models spur alarm abroad.
China rare-earth export curbs risk $6.5 trillion in production — Semafor
Why it matters: Rare-earth curbs threatening trillions in downstream production is the chokepoint that survives any chip strategy.
China's rare-earth export curbs put an estimated $6.5 trillion of global production at risk, the leverage point no semiconductor policy addresses.
Chinese memory giant CXMT oversubscribed 212 times in mega Shanghai IPO — Tech - South China Morning Post
Why it matters: A 212-times oversubscribed memory IPO is domestic capital pricing in Chinese chip self-sufficiency as inevitable.
Chinese memory maker CXMT's Shanghai IPO was oversubscribed 212 times, signalling domestic conviction in China's chip self-sufficiency push.
Chinese memory ban would cut off RAMpocalypse relief — www.theregister.com - Articles
Why it matters: Banning Chinese memory would tighten the very shortage it is meant to punish, which is the recurring flaw in chip containment.
A US ban on Chinese memory would cut off the main source of relief from the current RAM shortage, raising costs across the market.
Unitree Files for IPO, Aiming to Become 'First Embodied Intelligence Stock' — Pandaily - China Tech News, AI & Electric Vehicle Insights
Why it matters: Unitree listing as the first embodied-intelligence stock puts a public market price on Chinese humanoid capability.
Unitree filed for an IPO aiming to become the first listed embodied intelligence company.
Threat Intelligence (CTI)
[P1] Claude Code and DeepSeek Powered Chinese Cyber Espionage Campaign — Security Affairs
Why it matters: An American model driving the tooling and a Chinese open model doing the reasoning, inside a live government intrusion set — the arrangement defeats vendor-side disruption by design.
Hunt.io pivoted from known TencShell command-and-control infrastructure to 13 Hong Kong servers, one of which exposed an open directory of 2,431 files and 80 subdirectories containing victim source code, custom exploit scripts, cloned login pages and operator logs in Simplified Chinese, showing Claude Code as the execution engine handling agentic tool use, bash execution, session persistence and task parallelisation, with DeepSeek-v4-pro as the reasoning model driving attack logic, script generation and decision-making against government, Taiwanese industry and financial services targets.
severity high · exploited in the wild · EU: NIS2, CER Directive, AI Act
[P2] HelloNet campaign — new malicious modules launched through the ViPNet update system — Securelist
Why it matters: A Chinese-speaking group with persistent access inside the Russian state's own secure networking product — via its update mechanism.
Kaspersky found a malicious wtsapi32.dll in the ViPNet Update System directory, sideloaded by itcsrvup64.exe at Windows startup, delivering HelloInjector, HelloProxy, a Rust-based HelloBackdoor, HelloExecutor for command execution and HelloCleaner to erase ViPNet logs; the campaign began at least in May 2026, remains active, and hits Russian government, industrial, energy, transport, logistics and education targets.
severity high · exploited in the wild · EU: NIS2, CRA · actor Unknown Chinese-speaking group (30%)
[P2] Sandworm hackers have a CAPTCHA trick for Ukrainians — The Record from Recorded Future News
Why it matters: Russia's most destructive military intelligence unit is now using the same fake-CAPTCHA trick as commodity criminals.
CERT-UA observed Sandworm shifting initial access this spring and summer to ClickFix, serving fake CAPTCHA prompts on more than ten compromised websites during June and July that instruct Ukrainian visitors to paste a PowerShell command into Windows, which downloads malware enabling persistent access and later tooling; Sandworm also continues targeting Android devices with malware disguised as security apps distributed through messaging apps, collecting contacts, files, device information and real-time location.
severity high · exploited in the wild · EU: NIS2, CER Directive · actor Sandworm (APT44) (85%)
[P3] Two Scattered Spider Hackers Get 5.5 Years Each for £29 Million TfL Hack — The Hacker News
Why it matters: The first successful prosecution under the Computer Misuse Act's most serious section — for talking a help-desk into a password reset.
Owen Flowers, 18, and Thalha Jubair, 20, were each sentenced to five and a half years at Woolwich Crown Court on 16 July for the 2024 Transport for London intrusion, which left 148 TfL systems inoperable and forced all 27,000 employees to attend an office in person for password resets, at a cost the NCA and CPS put at £29 million; they bought partial credentials on criminal forums, reset 2FA on employee accounts, and socially engineered a help-desk worker into resetting a password.
severity medium · EU: NIS2, CER Directive · actor Scattered Spider (95%)
[P2] The Gentlemen Overtakes Qilin as Most Prolific Ransomware Threat — Infosecurity Magazine
Why it matters: The most prolific ransomware operation is now a Qilin defector paying affiliates 90% — and today's leak sites show it.
ReliaQuest reports The Gentlemen ran 300 incidents over three months against Qilin's 289, taking the top spot Qilin had held every quarter of the previous year, and claimed 121 victims in June alone; the group evolved out of a Qilin affiliate and recruits by paying 90% of collected ransoms against the usual 70-80%, packaged with a pre-built intrusion kit that lowers the skill floor for new operators.
severity high · exploited in the wild · EU: NIS2, GDPR · actor The Gentlemen (90%)
[P3] AI Has Enhanced Iran’s Asymmetric Playbook During the 2026 Conflict — Recorded Future
Why it matters: Six months of watching Iran fight with AI in the loop, and the conclusion is that AI multiplied the operations without being the source of the power.
Recorded Future assessed Iranian operations from January to June 2026 and found Tehran's hybrid model — asymmetric military operations, cyber, information warfare, proxy attacks and coercive state control — enhanced by AI as a force multiplier increasing speed, scale and effectiveness, including AI-enabled targeting and coordination claimed for the IRGC's Shahed-161 drone and AI-generated personas and imagery amplifying influence narratives across platforms.
severity medium · EU: CER Directive, NIS2, EU Hybrid Toolbox · actor Iran (state-directed hybrid operations) (75%)
Defence & National Security
A Humanoid Company Backed by Eric Trump Is Preparing Its Robots for War — WIRED
Why it matters: A politically connected humanoid firm preparing robots for war is where embodied AI, defence procurement and access converge.
A humanoid robotics company backed by Eric Trump is preparing its machines for military use.
Russia and China have a plan to disable Starlink, as Ukraine war comes to space (Ukraine Battlefield update: Day 1,604) — EUobserver
Why it matters: Joint Russian-Chinese planning against Starlink extends the Ukraine war into orbit against infrastructure the West now assumes.
Russia and China are developing plans to disable Starlink, extending the Ukraine war into the space layer.
Zelenskyy announces spy chief as new interim defense minister — Policy – POLITICO
Why it matters: Naming the intelligence chief as interim defence minister answers who replaces Fedorov, and what that signals about civilian control.
Zelenskyy named Ukraine's spy chief interim defence minister after Fedorov's dismissal triggered protests and allied criticism.
Intel Officials Predict the Pentagon’s Bill for the Iran War Will Exceed $100 Billion — WIRED
Why it matters: A hundred-billion-dollar war bill is the fiscal constraint that will shape every US procurement argument this year.
US intelligence officials predict the Pentagon's bill for the Iran war will exceed $100 billion.
German military to join French nuclear exercise for first time — Policy – POLITICO
Why it matters: German participation in a French nuclear exercise is a concrete step toward a European deterrent independent of Washington.
German forces will join a French nuclear exercise for the first time, a practical move toward European deterrence.
Poland indicts teen accused of performing sabotage for Russia — Policy – POLITICO
Why it matters: Prosecuting a teenager recruited online for Russian sabotage shows the recruitment pipeline is cheap, young and domestic.
Poland indicted a teenager accused of carrying out sabotage on behalf of Russia.
Digital Sovereignty & Identity
Proposed Europol reform dangerously erodes privacy, automates surveillance, and sidelines oversight — European Digital Rights (EDRi)
Why it matters: Expanding Europol's data powers while thinning oversight is the structural privacy fight that outlasts any single surveillance file.
EDRi warns the proposed Europol reform erodes privacy, automates surveillance and sidelines the oversight mechanisms meant to constrain it.
Aufrüstung der Geheimdienste: Innenministerium lässt Datenkäufe in der Grauzone — netzpolitik.org
Why it matters: A German ministry buying commercial data in an acknowledged legal grey zone is the purchase loophole that makes warrant requirements optional.
Germany's interior ministry is letting intelligence services buy commercial data in a legal grey zone, sidestepping the thresholds that govern collection.
Japan, EU validate cross-border digital wallet interoperability — Biometric Update
Why it matters: EUDI wallet interoperability reaching beyond the bloc is the first sign the European identity stack could become an exportable standard.
Japan and the EU validated cross-border interoperability between their digital identity wallets, extending the EUDI architecture past the bloc's borders.
Piero Cipollone: The cooperative spirit at the heart of the digital euro — ECB - European Central Bank
Why it matters: Cipollone framing the digital euro as cooperative infrastructure is the ECB answering the banks' capture objection directly.
ECB board member Piero Cipollone made the case for the digital euro as cooperative infrastructure shared between the central bank and intermediaries.
UK won’t ban VPNs, puts onus on platforms to prevent age check circumvention — Biometric Update
Why it matters: Declining to ban VPNs while making platforms liable for circumvention is the pragmatic limit of age-gating enforcement.
The UK will not ban VPNs, instead placing the burden on platforms to prevent users circumventing age checks.
Here’s the Truth About Whether Meta’s NameTag Face Recognition Tech ‘Exists’ — WIRED
Why it matters: Pinning down whether Meta's face recognition wearable actually exists matters because the denial itself shapes the regulatory clock.
WIRED examined the conflicting claims about whether Meta's NameTag face recognition technology exists, and what the ambiguity conceals.
Quantum & Cryptography
Fudan University Quantum Flash Technology Hits Theoretical Limit: Room-Temperature Single-Electron Storage Published in Science — Pandaily - China Tech News, AI & Electric Vehicle Insights
Why it matters: Room-temperature single-electron storage published in Science is a genuine physical milestone with long-run implications for memory and quantum control.
Fudan University reported room-temperature single-electron storage reaching the theoretical limit, published in Science.
Details of Alan Turing’s Voice Encryption System — Schneier on Security
Why it matters: The engineering detail of Turing's Delilah is a reminder that secure voice was solved once under far harder constraints than today's.
Newly surfaced details describe the design of Alan Turing's Delilah voice encryption system.
Cybersecurity & Threats
[P1] Security incident disclosure — July 2026 — Hugging Face - Blog
Why it matters: An autonomous agent framework running the entire intrusion at the hub hosting the world's open models is the first breach of its kind disclosed by the victim.
Hugging Face disclosed that an intruder chained two code-execution paths in its dataset processing — a remote-code dataset loader and a template injection in a dataset configuration — then used an autonomous AI agent framework to run many thousands of actions across a swarm of short-lived sandboxes, escalating to node-level access and lateral movement across internal clusters; limited internal datasets were accessed and several service credentials harvested.
severity high · exploited in the wild · EU: NIS2, GDPR, CRA, AI Act
[P1] CISA urges immediate action on actively exploited Fortinet flaws — BleepingComputer
Why it matters: Two unauthenticated command-execution flaws in a security appliance, exploited in the wild, with a three-day federal deadline.
CISA added two FortiSandbox flaws, CVE-2026-39808 and CVE-2026-25089, to the Known Exploited Vulnerabilities catalogue on 16 July, both rated 9.1 and both allowing unauthenticated attackers to execute operating system commands via crafted HTTP requests; federal civilian agencies must apply vendor mitigations by 19 July under BOD 26-04.
severity critical (CVSS 9.1) · exploited in the wild · CVE-2026-39808 · EU: NIS2, CRA
[P2] Coca-Cola Suspends US Fairlife Production Due to Ransomware Attack — SecurityWeek
Why it matters: Ransomware halting US production at a Coca-Cola dairy brand, confirmed in an 8-K — a named victim with physical consequences.
Coca-Cola disclosed in an 8-K filing that fairlife, LLC identified unauthorised third-party access to a portion of its systems including production-related systems in a ransomware event, temporarily suspending all US production; Canadian production is unaffected and the company says product quality and safety were not impacted.
severity high · exploited in the wild · EU: NIS2, GDPR
[P2] ‘ClickLock Stealer’ Bypasses macOS Security With Social Engineering, Process Killing — SecurityWeek
Why it matters: Malware that kills the desktop every 210 milliseconds until the victim types their password is coercion engineered as a feature.
ClickLock Stealer arrives as a command pasted into Terminal and asks for the login password behind a fake dialog; when the victim cancels, it installs two LaunchAgents and exits, and at next login kills Finder, the Dock, Spotlight, Terminal, Activity Monitor and major browsers every 210 milliseconds for up to 83 hours, leaving a single password box on a dead desktop before harvesting the Keychain, browser credentials and password manager extensions.
severity high · exploited in the wild · EU: GDPR, NIS2
[P3] Single Prompt Enables ChatGPT to Execute Full Cyber-Attack Chain, Researchers Claim — Infosecurity Magazine
Why it matters: A frontier model taking a network to domain admin in forty minutes from one prompt — significant, and a lab result, which matters.
Cato Networks researchers gave ChatGPT-5.5 a single high-level objective in a controlled Active Directory environment built to resemble a typical enterprise, and the agent planned and executed the full lifecycle — reconnaissance, exploitation, internal discovery, privilege escalation, lateral movement and exfiltration — reaching admin-level privileges in roughly 40 minutes, adjusting its approach from observation rather than following a fixed sequence.
severity medium · EU: AI Act, NIS2
[P3] Security researchers find stalkers abusing Chrome’s sync feature — CyberScoop
Why it matters: Stalkerware without the malware — a browser feature doing the surveillance, with nothing to detect and no warning shown.
Certo reports a growing number of cases in which stalkers briefly unlock a victim's phone, tablet or computer, add an attacker-controlled Google account to Chrome and enable sync, thereafter receiving the victim's browsing history and saved passwords remotely with no warning ever displayed to the victim.
severity medium · exploited in the wild · EU: GDPR, Istanbul Convention