Poisoned Emails Root Cisco Gateways as PAN-OS Zero-Day Hits Firewalls

AI & Machine Learning

Ricursive Intelligence, the chip-design startup founded by former Google AlphaChip researchers Anna Goldie and Azalia Mirhoseini, used the TechCrunch Disrupt stage on 14 October to argue that AI-designed silicon is part of the answer to the compute crunch. Rather than splitting a chip into dozens of separately optimized blocks, Ricursive’s system optimizes whole chips and learns from each iteration, sometimes producing unconventional wave-like layouts. Goldie said compute demand could grow at least 10,000x over five years and called that estimate potentially conservative. The founders claim design cycles could shrink from months or years to about a week, which would make custom chips for niches such as space, AR/VR and biological computing financially viable. The pitch puts the company, which raised $335M at a $4B valuation earlier this year, squarely in the efficiency-over-more-datacenters camp. Source: TechCrunch Verified: True

Linux kernel maintainers have agreed to deprecate the “Assisted-by” tag that required contributors to disclose when an AI coding tool helped produce a patch, Nvidia kernel developer Sasha Levin told the mailing list after the Maintainers Summit reached consensus. The rule, proposed in July 2025, entered the kernel documentation only in January 2026. Red Hat’s Jeff Layton cited three reasons for the reversal: compliance was patchy, the tag worked as free advertising for proprietary LLM vendors, and nobody could say why the data was needed. Greg Kroah-Hartman said LLMs are now good enough for kernel work, while Linus Torvalds has long argued contributors must understand their code regardless of tooling. Some developers, including Laurent Pinchart, objected and want an alternative disclosure mechanism, so the debate over AI provenance in open source is far from settled. Source: The Register Verified: True

Consumer Hardware

Deezer says it will add a toggle before the end of 2026 that hides AI-generated tracks, albums and related artists from search results. The streaming service reports that AI music has exploded on its platform: AI tracks made up about 10% of daily uploads when it introduced detection tools in 2025, but by July 2026 they accounted for more than half, roughly 90,000 tracks a day. A 2025 Deezer-Ipsos survey found around 45% of its customers wanted a way to remove AI music from their feeds. The move arrives the same day YouTube Music introduced new labelling for AI-generated songs, signalling that streaming platforms are now competing on provenance controls rather than just catalogue size. Source: TechCrunch Verified: True

Hiyd, a Startup Battlefield top-20 company at TechCrunch Disrupt, unveiled a 2.7 mm, 7-gram health-tracking disc that clings to the back of an existing watch using a textured microsuction layer instead of magnets or adhesive. The sensor tracks heart rate, heart rate variability, sleep stages, movement, calories and VO2 max estimates, syncing to iOS and Android apps that use AI to analyse activity. The team, led by former Apple manufacturing design engineer Salah Batroukh, plans to add nutrition tracking focused on how eating habits affect sleep and workouts. Hiyd will launch a Kickstarter in November and says more than 16,000 people have already placed deposits, a sign of appetite for health tracking that does not require replacing a beloved analogue watch. Source: TechCrunch Verified: True

Cybersecurity

Cisco has patched CVE-2026-76519, a CVSS 9.8 zero-day in its Secure Email Gateway that lets an unauthenticated attacker write files and execute commands as root simply by sending a malicious email. The bug sits in the Advanced Malware Protection feature’s attachment handling and affects physical and virtual appliances with File Reputation enabled and Content Scanner Tools older than 25.3.0.5525. Cisco and CISA both confirmed exploitation before the fix shipped, and CISA added the flaw to its Known Exploited Vulnerabilities catalog with a weekend deadline for federal agencies. There are no workarounds, and it is the second root-via-email zero-day in Cisco’s gateway in roughly two months after CVE-2026-76461 was abused in September. Source: The Register Verified: True

Palo Alto Networks warned that CVE-2026-0322, a high-severity (8.6) remote code execution flaw in the PAN-OS management interface, has been exploited as a zero-day in limited attacks. Attackers with valid credentials and network access to the management interface can run code as root on PA-Series and VM-Series firewalls and Panorama appliances, while Cloud NGFW and Prisma Access are unaffected. Fixes are available across PAN-OS 10.2 through 12.2, with one remaining build due on 15 October, and the company published three attacker IP addresses. No workaround exists beyond restricting management access to trusted internal IPs, and CISA has ordered federal agencies to mitigate by 17 October. Source: BleepingComputer Verified: True

Microsoft, Japan’s National Police Agency and partners including Health-ISAC, Cloudflare, Oracle, Proofpoint, Resecurity and TRM Labs disrupted a Chinese-language cybercrime marketplace behind more than one billion phishing messages sent to over 75,000 organisations in about 150 countries. A Florida federal court order let Microsoft and Health-ISAC seize domains tied to the YAM Store marketplace (Storm-3119) and the CoGUI, Haozi and Iron Man phishing kits (Storm-2990), the last of which offered more than 2,300 impersonation templates. Over 750 million of the messages targeted Microsoft customers this year alone. Operators revived expired domains of real organisations, used AI to fabricate fake companies, and leaned on device-code phishing to steal tokens; Microsoft deactivated 20,000 linked accounts and blocked 100,000 domains. Source: BleepingComputer Verified: True

Federal prosecutors have charged OpenAI engineer Abdulrahman Alfozan with attempting to intercept oral communications after a Seeed XIAO ESP32-S3 Sense recording device was found wrapped in toilet paper and plugged into a TV in OpenAI’s executive conference room. Staff discovered the device on 30 September, and a second recorder turned up in a bag he left in a printer room on 5 October, according to the complaint. Investigators allege he queried an internal productivity agent about meeting rooms and calendars and that an AI chatbot account linked to his spouse was asked how to build and modify an audio recorder. OpenAI says it has no indication that model weights, source code or customer data were taken and that he has been terminated; the charge carries up to five years in prison. Source: The Register Verified: True

Enterprise Infrastructure

Cisco and Nvidia unveiled the Ultra-Dense Small Form Factor Pluggable (USFP) at the Open Compute Project Global Summit, an OSFP-sized optical module designed to carry up to 12.8 Tbps of Ethernet using as many as 32 lanes of 400 Gbps. That is four to eight times today’s top pluggables, though each module can draw up to 100 watts versus roughly 20 watts now, and Cisco’s initial spec reportedly starts at 16 lanes, or about 6.4 Tbps. A single USFP could replace the cabling for eight GPUs, and a future 204.8 Tbps switch would need just 16 ports to feed 128 endpoints. The companies intend it as an open standard aimed at dense AI racks such as Nvidia’s Vera Rubin Ultra, while Cisco also confirmed Microsoft is using its 51.2 Tbps Silicon One P200 for multi-datacenter “scale across” links. Source: The Register Verified: True

Gartner warns that AI capital write-downs could exceed $4 trillion by 2032 as infrastructure spending runs far ahead of revenue. AI providers are on track to spend nearly $1 trillion on infrastructure in 2026 and a further $6 trillion between 2027 and 2030, according to the report. To cover that capex and financing at a 12% return on invested capital, AI datacenter operators would need about $8.5 trillion in revenue, but Gartner forecasts no more than $3.6 trillion from AI services they host. The firm expects 80% of neocloud GPU-as-a-service providers not to survive as independent companies beyond 2030, while diversified hyperscalers should weather the consolidation better. Source: The Register Verified: True

Policy & Regulation

CCIA Europe, the lobby group whose members include Amazon, Google and Meta, released an ECIPE study claiming the European Commission’s proposed Cloud and AI Development Act could cost €115 billion over a decade, more than six times the Commission’s own upper estimate of €18 billion. The study puts public-sector costs at €33.3 billion and private-sector costs at €80 billion, assuming sovereignty requirements extend to 25% of cloud workloads in essential industries such as banking, healthcare and energy. Code rewriting and re-architecting account for 69% of the projected bill, and the group is targeting Article 31 of the draft law. CCIA wants rules limited to sensitive workloads and based on technical safeguards rather than corporate origin, even as its members market their own European sovereign clouds. Source: The Register Verified: True

Flock Safety CEO Garrett Langley used a TechCrunch Disrupt panel to defend the company’s license-plate and surveillance cameras amid a national backlash that has cost it contracts and forced layoffs. Langley argued society should not “pause progress” if it wants accountability for crime, likening Flock to driverless cars where individual failures warrant fixes rather than shutdowns. He conceded the company is “far from perfect” but said Flock would offer cities recommendations and defaults rather than mandating safeguards, and blamed some cancellations on misinformation. Pressed on how abuse would be prevented, he offered no concrete measures, a gap likely to fuel local and federal efforts to restrict the technology after recent court criticism and congressional proposals. Source: TechCrunch Verified: True