Thursday 25 June 2026

OpenAI's "Jalapeño" Chip Is the Moment It Stops Renting Its Own Brain

OpenAI's first custom silicon, a macOS security gap anyone can exploit, and a coordinated takedown of two major malware-as-a-service operations — plus Australia's health sites caught covertly tracking patients.

Lead story

OpenAI's "Jalapeño" Chip Is the Moment It Stops Renting Its Own Brain

For years, OpenAI has been paying Nvidia rent on its own ambitions. That changes today. The company has unveiled Jalapeño, its first custom inference chip, built in partnership with Broadcom — and it's a signal that the AI infrastructure war has moved from software to silicon.

What it is. Jalapeño is an ASIC — an application-specific integrated circuit — optimised specifically for running large language model inference, the process of generating responses to prompts. It's not a training chip. OpenAI isn't trying to replace the H100s it uses to teach its models; it's trying to dramatically cut the cost of serving the answers those models produce. Every token you generate costs OpenAI money, and right now that money largely flows to Nvidia.

Why it matters. Inference is where AI companies live or die commercially. Training happens once (or a few times). Inference happens billions of times a day. By owning that layer of the stack, OpenAI can lower its per-token cost, pass some of that on to customers, and — crucially — stop being structurally dependent on a hardware supplier whose chips are also sold to every competitor OpenAI has. Think of it like a restaurant that has finally stopped renting its kitchen from the same company supplying everyone else on the street.

The Broadcom angle. Choosing Broadcom as the manufacturing partner is notable. Broadcom has become the go-to custom chip partner for hyperscalers — Google's TPUs are a Broadcom collaboration. OpenAI is essentially following the playbook Google wrote a decade ago. The difference is that Google had armies of chip engineers in-house before it went down this path. OpenAI is moving faster and leaning harder on the partnership model.

What it doesn't tell us. OpenAI hasn't released performance benchmarks, hasn't said when Jalapeño will be powering production traffic at scale, and hasn't disclosed whether this changes its Nvidia procurement plans in the near term. The announcement is strategic positioning as much as it is a product launch. Expect the gap between "unveiled" and "widely deployed" to be measured in quarters, not weeks.

The broader picture. This is the third major AI lab to go custom silicon in the last 18 months. Google has TPUs, Amazon has Trainium and Inferentia, and now OpenAI has Jalapeño. Microsoft — OpenAI's biggest backer and cloud partner — has its own Maia chip. The implication for Nvidia is not immediate pain, but it's structural pressure. The hyperscaler custom silicon playbook reduces Nvidia's total addressable market over time by carving out the most predictable, highest-volume inference workloads.

Watch for: whether OpenAI's API pricing drops in the next two quarters, which would be the clearest signal that Jalapeño is doing real work. If prices stay flat, the chip is either still in testing or the cost savings are being reinvested in capacity rather than passed on.

Also today

Operation Endgame Strikes Again: Amadey and StealC Infrastructure Smashed

Microsoft, Europol, and a coalition of cybersecurity firms including Bitdefender, ESET, and Bitsight have dismantled more than 300 command-and-control servers powering the Amadey and StealC malware operations. The two tools have been used in tandem — Amadey as a loader, StealC as a credential-harvesting infostealer — feeding ransomware gangs with initial access and stolen data. Around 27 million compromised credentials were recovered. Notably, Microsoft used AI analysis to link the two operations in a civil racketeering lawsuit, and this marks the first time a single court action has targeted two distinct cybercrime tools simultaneously. The operation is an extension of the Operation Endgame campaign that began disrupting cybercrime infrastructure in 2024.

CyberScoop ↗

Cordyceps: The CI/CD Flaw Class That Could Hijack Millions of Repos

Researchers at Novee Security have disclosed a new class of CI/CD workflow vulnerability they've named Cordyceps — after the fungus that takes over insect brains, which feels apt. The flaws allow unauthenticated attackers to hijack GitHub Actions workflows and inject malicious code into open-source repositories. Over 300 repositories at organisations including Microsoft, Google, and Apache were found to be vulnerable. Unlike a single CVE, Cordyceps describes a repeating structural pattern in how pipelines are configured — meaning it's a category of mistake, not a one-off bug. With open-source supply chain security already under intense scrutiny following last year's npm poisoning campaigns, this adds another layer of systemic risk for any organisation pulling dependencies from GitHub-hosted projects, including Australian government and enterprise users.

The Hacker News ↗

macOS Gap Lets Standard Users Silently Kill Security Tools

Security researchers have detailed a chained attack against macOS that allows a standard, non-administrator user account to silently disable endpoint security agents — including integrated browser tools — without kernel exploits or elevated privileges. The technique abuses legitimate macOS behaviours rather than exploiting a traditional vulnerability, which makes it harder to patch cleanly and means existing agent self-protection mechanisms don't catch it. Apple hasn't issued a fix as of publication. For organisations running macOS fleets with endpoint detection tools — a growing cohort in Australian professional services and financial firms — this is a meaningful gap: an attacker who gets a foothold as a low-privilege user can effectively go dark before the real work begins.

SecurityWeek ↗

CISA Flags Max-Severity Ubiquiti and Lantronix Flaws Under Active Exploitation

CISA has added critical vulnerabilities in Ubiquiti's UniFi OS and Lantronix EDS5000-series serial-to-ethernet servers to its Known Exploited Vulnerabilities catalogue. The Lantronix flaw, CVE-2025-67038, carries a CVSS score of 9.8 and allows unauthenticated remote code execution. The Ubiquiti bugs enable attackers to make system-level changes and inject commands without credentials. Both product lines are common in enterprise and industrial environments — Ubiquiti gear in particular is ubiquitous in small-to-medium business networks, co-working spaces, and regional organisations across Australia. Federal agencies have until 26 June to patch. Everyone else should treat that deadline as their own.

Bleeping Computer ↗

Cisco Unified CM Zero-Day Exploited After PoC Drops

A high-severity server-side request forgery vulnerability in Cisco Unified Communications Manager (CVE-2026-20230) is now being actively exploited, days after a proof-of-concept demonstrated a file-write path to root-level access. Mandiant separately detailed a related Cisco SD-WAN zero-day that gave an unnamed attacker the highest possible access level at a communications service provider — and it's unclear whether that actor gained visibility into internal traffic. Cisco Unified CM is widely deployed in enterprise telephony environments. Australian organisations running on-premise Cisco comms infrastructure should treat patching as urgent, and review network logs for anomalous HTTP requests to the management interface.

The Hacker News ↗

Australia's Health Sites Were Quietly Tracking Patients — OAIC Found Out

The Office of the Australian Information Commissioner has completed a sweep of Australian health sector websites and found that JavaScript tracking pixels on multiple sites were transmitting sensitive patient information — including health condition indicators and appointment details — to overseas advertising platforms. The data was flowing without users' knowledge or meaningful consent. Health information is among the most sensitive categories under the Privacy Act, and the OAIC's findings are likely to trigger formal investigations. This is a textbook example of the ad-tech stack running faster than governance: someone installed a marketing pixel for analytics purposes, and nobody noticed it was also siphoning health data to a US ad platform.

iTnews ↗

Anthropic's Mythos Model Found Vulnerabilities in Classified US Government Systems

Anthropic's Mythos AI model — built specifically for security research tasks — has been used to identify real vulnerabilities in classified US government systems, according to a report. Details remain thin: it's unclear whether the flaws were immediately exploitable or how many systems were affected, but the fact that an AI model is now doing red-team work inside classified environments is a significant threshold. This is the flip side of the AI-cyber risk coin: the same capability that makes AI dangerous in adversaries' hands is being pointed defensively at US infrastructure. For Australian readers, the ASD's own AI-assisted vulnerability research programme is worth watching for similar disclosures.

iTnews ↗

Gemini 3.5 Flash Gets Computer Use — Google's Answer to Claude's Operator Mode

Google DeepMind has introduced computer-use capabilities in Gemini 3.5 Flash, allowing the model to interact directly with on-screen interfaces — clicking buttons, filling forms, navigating applications — without needing a dedicated API. This puts Gemini squarely in competition with Anthropic's Claude, which pioneered the computer-use category for commercial models. The significance is in the model chosen: Flash is Google's fast, cheap inference model, meaning computer-use is now accessible at a price point that makes high-volume agentic automation practical. Expect rapid uptake in RPA-adjacent workflows across enterprise software. The attack surface implications for agentic AI — prompt injection via screen content, for instance — remain an open and underexplored problem.

Google DeepMind ↗

The Tokenpocalypse: Enterprises Are Now Rationing AI Usage

Leaked Accenture audio and reporting from multiple outlets confirm what finance teams have been quietly discovering: enterprise AI costs are exploding, driven not by complex workloads but by mundane ones — employees using frontier models to convert PDFs to slide decks, summarise short emails, or reformat spreadsheets. Companies are now imposing token budgets, tiering access, and routing low-complexity tasks to cheaper models. The period of unlimited AI experimentation inside large organisations appears to be ending. For CIOs, this is actually a useful forcing function: it requires a proper taxonomy of tasks and a deliberate model-selection strategy rather than defaulting to the most capable (and expensive) model for everything.

TechCrunch AI ↗

White House Accelerates Post-Quantum Cryptography Deadline

A new US executive order has significantly brought forward the deadline for federal agencies — and by extension, federal contractors — to migrate off quantum-vulnerable cryptographic algorithms. The order cites national security risk if post-quantum cryptography isn't adopted before cryptographically relevant quantum computers emerge. NIST finalised its post-quantum standards in 2024, so the algorithms exist; the challenge is the sprawling legacy infrastructure that still runs RSA and elliptic-curve crypto. Australia's own post-quantum migration guidance from the ASD has been relatively measured in tone — this US order may prompt a rethink of the urgency framing, particularly for Australian organisations with US government contracting relationships.

Ars Technica ↗

KDDI Exposes 14.2 Million Email Credentials Across Five ISPs

Japanese telecoms giant KDDI has disclosed a data exposure affecting approximately 14.2 million managed email accounts across five internet service providers it operates. The incident involved credentials — usernames and passwords — sitting accessible in a way that left customers of those ISPs at risk. Japan's telecommunications sector is tightly integrated with Australian enterprise networks through regional connectivity infrastructure, and KDDI has a significant wholesale and B2B presence in the Asia-Pacific. The scale — 14 million accounts — puts this in the same tier as Australia's Optus breach in terms of raw credential exposure, though the full downstream impact depends on how many affected users reuse passwords elsewhere.

The Register ↗

Sources consulted