Thursday 13 August 2026

AI Is Finding Bugs Faster Than Humans Can Patch Them — and Microsoft's Patch Tuesday Is the Proof

Microsoft's Patch Tuesday has ballooned to nearly 400 fixes a month — and AI-assisted bug discovery is why.

Lead story

AI Is Finding Bugs Faster Than Humans Can Patch Them — and Microsoft's Patch Tuesday Is the Proof

Microsoft shipped fixes for 398 security vulnerabilities this month. To put that in perspective: a typical Patch Tuesday before AI-assisted vulnerability discovery became mainstream sat around 80 bugs. That's roughly five times the volume, and it's not because Microsoft's code suddenly got five times worse.

The culprit — or the credit, depending on how you look at it — is AI. According to reporting from The Record, the explosion in patch volume tracks almost precisely with the widespread adoption of AI-assisted bug-hunting tools. Researchers are now finding and reporting vulnerabilities faster than the industry's patching workflows were ever designed to absorb.

What that means in practice is a prioritisation crisis. Security teams can't treat 400 CVEs the same way they once treated 80. Most of August's bugs aren't critical, but buried in the pile are the ones that genuinely matter: one actively exploited zero-day, two publicly detailed before patches dropped, and a fresh Windows flaw already weaponised by North Korea's Lazarus Group (more on that below). The signal is getting harder to find in the noise.

The Lazarus angle is worth pausing on. Check Point Research linked the zero-day — patched this month — to Operation Dream Job, a long-running North Korean campaign that lures targets in defence and aerospace with fake job offers before deploying malware. The new backdoor, dubbed ForestTiger, has been used against targets in France, Germany, Brazil, and India. CISA has given US federal agencies two weeks to patch.

The deeper issue is structural. The security industry built its patching culture around a monthly cadence — a rhythm that made sense when bugs were found by humans working at human speed. AI doesn't work on that schedule. It can fuzz, analyse, and chain vulnerabilities in minutes. The gap between "bug found" and "patch shipped" hasn't shrunk to match; if anything the queue is lengthening.

There's also a less-discussed flip side. If defenders are using AI to find bugs faster, so are attackers. The same tools that produced this month's 398-fix avalanche are available to anyone. Microsoft's patch volume isn't just a workload problem for sysadmins — it's a measure of an arms race that's visibly accelerating.

For Australian organisations, the volume problem is acute. The Australian Cyber Security Centre has repeatedly flagged patch prioritisation as a core failure point in local breaches. With Patch Tuesday now resembling a monthly firehose, the Essential Eight's patching controls — which require critical patches applied within 48 hours for the highest maturity tiers — are going to demand more automation, not just more attention.

The practical advice hasn't changed, but the stakes have: patch the actively exploited bug first (CVE tracking is your friend), treat publicly disclosed-but-unpatched flaws as near-critical, and treat the rest as a triage exercise rather than a to-do list. If you're still doing this manually, you're already behind.

What to watch: whether the broader industry responds with faster patch cycles, better automated deployment tooling, or — most likely — a quiet recalibration of what "prompt patching" actually means when the volume triples again next year.

Also today

Lazarus Group Deploys ForestTiger Backdoor via Windows Zero-Day

North Korea's Lazarus Group has been caught exploiting a now-patched Windows zero-day as part of its long-running Operation Dream Job campaign. The group used fake job postings to lure targets in the defence and aerospace sectors across France, Germany, Brazil, and India before deploying a previously unseen backdoor called ForestTiger. Check Point Research attributed the campaign to Lazarus after reverse-engineering the implant. CISA has ordered US federal agencies to patch within two weeks. Australian defence contractors and aerospace firms should note: Operation Dream Job has previously cast a wide net, and similar lure campaigns have been observed targeting AUKUS-adjacent organisations.

The Hacker News ↗

AI API Flaw Exposed Reasoning Secrets Across OpenAI, Anthropic, and Google

Researchers have disclosed a significant flaw in the way all three major AI API providers — OpenAI, Anthropic, and Google — handled encrypted reasoning objects passed between sessions. A reasoning block generated in one session could be replayed into a different session and, during testing, leaked internal chain-of-thought content including API keys and passwords. The flaw has since been addressed, but it raises pointed questions about the security architecture of AI reasoning APIs, which are now embedded in a growing number of enterprise applications. Any organisation using these APIs to process sensitive data should review whether session isolation was properly enforced during the affected window.

The Hacker News ↗

Coin-Sized Device Can Redirect a Boeing 737's Autopilot in Under 60 Seconds

Security researchers disclosed a sobering aviation finding at this week's conferences: a device small enough to fit in a pocket can be plugged into an external hatch on a Boeing 737 and used to redirect the autopilot or corrupt the flight plan — all in under a minute. The researchers were careful to note this requires physical access to the aircraft's exterior, but the ease of exploitation once that access is gained is the alarming part. Avionics security has long been treated as an air-gapped non-problem; that assumption is looking shakier. Given Australian carriers operate 737 variants extensively, this is one to watch for regulatory follow-up from CASA.

WIRED Security ↗

VMware vCenter Flaw Under Active Exploitation — Patch Now

Attackers have begun actively exploiting CVE-2026-59310, a critical (CVSS 9.8) directory traversal vulnerability in VMware vCenter Server that allows unauthenticated remote code execution. Patches were available before exploitation began, making this a straightforward but painful case of patch lag. vCenter sits at the heart of most enterprise VMware deployments — compromise it and you effectively own the virtualised environment beneath it. This is the kind of flaw ransomware operators historically weaponise within days of PoC availability. Australian enterprises running vSphere environments should treat this as an emergency patch, not a scheduled maintenance item.

The Hacker News ↗

LiteLLM Supply Chain Attack Hit 2,500+ Organisations

Two malicious releases of LiteLLM — a popular Python library that provides a unified interface to multiple AI APIs — sat on PyPI for roughly 40 minutes in March before being pulled. That was long enough. Threat intelligence firm CloudSEK says the attackers captured around 434,000 files from systems that installed the compromised versions, with potential exposure now mapped to more than 2,500 organisations. Stolen data included cloud credentials, SSH keys, Kubernetes tokens, and database passwords. LiteLLM is widely used by AI developers, making this a high-value supply chain hit. Any team that was running automated dependency updates in March should audit their credential stores.

SecurityWeek ↗

Adobe Drops Three CVSS 10.0 Patches for ColdFusion and Campaign Classic

Adobe has patched three maximum-severity vulnerabilities — all rated CVSS 10.0 — across ColdFusion, Commerce, and Campaign Classic. The most serious, CVE-2026-48362, is an OS command injection flaw in ColdFusion that allows unauthenticated remote code execution. ColdFusion has a long and unhappy history of being targeted by both financially motivated attackers and nation-state actors; unpatched public-facing instances tend to get found quickly. Australian government agencies and enterprises still running ColdFusion for legacy web applications — and there are more than you'd think — should treat this as an urgent out-of-cycle patch.

The Hacker News ↗

UK Criminal Records Office Breached Three Times Over Two Years — Nobody Noticed

Britain's ACRO Criminal Records Office was compromised in three separate intrusions over a two-year period, none of which were detected in real time. The cause is almost comically mundane: antivirus alerts went unread, and the organisation's content management system was never patched. A formal reprimand notice confirms ACRO still cannot determine whether sensitive criminal records data was actually exfiltrated. It's a case study in security hygiene failure at exactly the kind of agency that holds highly sensitive personal data. Australian equivalents — state and federal criminal records bodies — operate under similar constraints; the OAIC's guidance on mandatory breach reporting would apply if a comparable failure occurred here.

The Record ↗

Prompt Injections Are Being Used as Honeypots to Trap AI Attackers

Researchers from Tracebit have found that embedding prompt injection text alongside sensitive credentials stored in AWS — API keys, passwords, cryptographic material — is surprisingly effective at neutralising AI-driven attack agents. When an LLM-based attacker encounters the injected prompt, it triggers the model's safety guardrails, causing it to abort the operation entirely. It's a defensive judo move: using the attacker's own model against itself. Bruce Schneier flagged the research as genuinely novel. The technique won't work forever — attackers will eventually tune their models to ignore such prompts — but it's an interesting asymmetry while it lasts.

Schneier on Security ↗

Spectre Haunts RISC-V: Speculative Execution Attacks Still Very Much Alive

Eight years after Spectre first broke the internet's collective calm about CPU security, researchers have demonstrated that several RISC-V chips are susceptible to the same class of speculative execution side-channel attacks. RISC-V's open architecture was partly pitched as a path to cleaner, more auditable hardware — but the underlying microarchitectural trade-offs that made Spectre possible haven't gone away just because the instruction set changed. The finding matters as RISC-V gains traction in embedded systems, edge devices, and increasingly in data centre silicon. Mitigations will again carry a performance penalty, the size of which varies by implementation.

The Register ↗

Twitch Has Been Training Amazon's AI on Streams for Years

Twitch has confirmed that user-generated content on its platform "may be used for future generative AI model improvements" — and has been for years — quietly rolling this into its terms of service without making it prominently obvious to creators. Amazon, which owns Twitch, has now added an opt-out setting, but it's buried in account preferences. The disclosure sits alongside a broader industry pattern of platforms quietly monetising user content for AI training before opting-out mechanisms exist. Australia's Privacy Act reforms, currently before parliament, include stronger provisions around secondary use of personal data that would likely require more prominent disclosure of this kind of practice.

Ars Technica ↗

Google Announces Pixel 11, Watch 5, and a New Tracking Tag

Google has unveiled its 2026 hardware lineup: the Pixel 11 smartphone, Pixel Watch 5, and a new Pixel Tag — its answer to Apple's AirTag. The emphasis this year is less on raw specs and more on AI-native experiences baked into the hardware, with Google leaning hard on on-device Gemini integration across the range. The Pixel Tag is notable given the regulatory scrutiny tracking devices have faced globally; Google says it has implemented the cross-platform unwanted tracking alerts that Apple and Google jointly standardised last year. Australian availability and pricing haven't been confirmed but Google has maintained strong local distribution for recent Pixel generations.

Ars Technica ↗

737 Fake Chrome VPN Extensions Were Routing Traffic Through Attacker Infrastructure

Researchers have identified 737 malicious Chrome extensions — mostly posing as free VPN and proxy tools — that were intercepting browser traffic and routing it through attacker-controlled proxy infrastructure. The extensions were published across at least 40 developer accounts and clocked up over 75,000 installs, with Russian-speaking users as the primary target. Around 274 of the extensions were found to be impersonating 66 legitimate VPN brands. The Chrome Web Store's extension vetting process has faced recurring criticism; this campaign appears to have run for an extended period before detection. Users relying on free VPN extensions for any sensitive browsing should audit what's installed.

The Hacker News ↗

Sources consulted