The year's best defence against stolen logins shipped switched off

Google has added the strongest protection yet against a form of account theft that has quietly grown as everything else got safer [1]. Then it left it off for nearly everyone.

That is not a scandal. It is testing, and testing is sensible. But it is also the week in one line. The defences exist. The settings exist. What decides who gets robbed is whether somebody flipped a switch.

The best answer yet, and almost nobody has it

The attack Chrome is now defending against is session-cookie theft [1]. A session cookie is the small token a website stores in your browser after you log in. It saves you re-entering your password on every page. It also proves, to that site, that you are already you.

That token has become the prize. Two-factor codes and passkeys made stealing a password much less useful [1]. So attackers stopped chasing passwords and started stealing the proof-of-login instead, using malware that scrapes browsers and fake login pages that sit in the middle of a real sign-in [1]. Paste the stolen cookie into your own browser and you get the same access the owner has.

Chrome’s answer is called device-bound session credentials [1]. The browser generates a secret key and stores it inside a small locked chip in your machine. On Windows that chip is the TPM; on Macs and iPhones it is the secure enclave [1]. The website then asks the browser to sign each session with that key. The chip will hand back an answer, but never the key itself.

“The attacker can steal the cookie, but they can’t answer a DBSC challenge,” researcher Scott Helme told Ars Technica [1]. The stolen token becomes useless.

The catch is the rollout. It works only in recent Chrome releases on Windows and Mac, and even there it is switched on for a limited set of users while Google tests it [1]. It is unclear when other Chrome-based browsers will follow [1].

Passkeys are not stored where most people assume

The same week produced three separate pieces of research defeating passkeys, none of which broke the underlying maths [3].

Passkeys replace passwords with a key pair split between your device and the site. The common belief is that the private half lives in that same locked chip. It usually does not. The FIDO standards do not require passkeys to be kept in dedicated hardware, and most platforms and passkey apps do not put them there [2].

That default is what one attack used. Palo Alto researcher Arie Olshtein showed that on a Windows machine already infected with malware, his technique could pull every passkey stored by the Google Password Manager app [2]. Separately, SpecterOps demonstrated a chain that impersonated privileged Microsoft accounts while still satisfying phishing-resistant two-factor checks [3]. An independent researcher showed malware in a signed-in Windows session using a hardware-backed Windows Hello key without prompting the user again [3].

Microsoft issued a fix for one piece of this, CVE-2026-34348, a flaw in the Windows Event Logging Service rated 6.5 out of 10 [3]. A CVE is just a public catalogue number for a specific flaw. Microsoft says it has also applied mitigations for the passkey issue SpecterOps reported [3].

Passkeys are still far better than passwords. The point is narrower: where the key sits is a configuration choice, and the choice most software makes for you is software storage.

Around four hundred fixes, and one already in use

Microsoft’s monthly patch release on 11 August was enormous. SecurityWeek and ZDNet counted 421 flaws fixed [4][6]; Brian Krebs counted at least 398, of which 42 were rated critical [5]. The gap is a counting convention, not a dispute. Krebs notes it is double June’s then-record batch, and that Microsoft attributes the surge to bugs now being found with the help of AI [5].

One flaw was already being used. CVE-2026-68820 sits in afd.sys, a core Windows networking driver [4][5]. It does not get an attacker in; it lets one who is already in become the administrator [5]. “This isn’t a front-door bug,” Automox’s Landon Miles wrote. “It’s step two in a chain” [5].

North Korea’s Lazarus group had been using it since early July, against defence, aerospace and aviation firms in Europe and India, approached through fake job offers [7]. Check Point found the exploit built into a new version of the group’s kernel-level rootkit, which switches off security software’s reporting [7].

Elsewhere: Metabase patched a critical flaw already being exploited that handed unauthenticated attackers admin access to its analytics tool [16]. GeoServer has an unpatched flaw being probed within hours of disclosure, with hundreds of attempts logged and no fix yet available [17]. CISA, the US cyber-defence agency, told federal agencies to patch a Progress LoadMaster flaw scored 9.6 out of 10 that has been exploited since late June [18].

The alarm that worked, and nobody read

Britain’s criminal records office, ACRO, was reprimanded by the data regulator this week over a breach nobody caught for seven months [8].

An intruder had access to ACRO’s website and its Kentico content management system from August 2022 to March 2023, exposing data on 10,920 people [8]. That data included names, dates of birth, National Insurance numbers, passport and driving licence details, bank details and criminal-record information [8]. Complaints came from people applying for child protection certificates and from domestic-violence victims [8].

The Information Commissioner’s Office found two failures, and both are about who was configured to act. ACRO’s outsourced IT provider patched the operating system but not the Kentico software [8]. The web supplier applied CMS patches, but was not responsible for spotting when they were needed [8]. ACRO itself did not monitor for them at all [8]. Nobody owned the gap.

The second failure is sharper. ACRO had security software installed, and it generated alerts. Those alerts were “not reviewed or acted upon” [8]. Had they been investigated at the time, the regulator said, further activity could likely have been prevented [8]. The intrusion was only discovered because ACRO happened to be investigating a different one, and it still cannot say whether the data was actually taken [9]. The regulator did note one thing ACRO had done right: the network was segmented, which limited how far the intruder could reach [8].

Attacks built to walk around the settings

Two supply-chain stories this week are about protections being routed around rather than broken.

The compromise that exposed credentials at more than 2,500 organisations turned out to trace back further than first reported [11]. SOCRadar found that over 95% of affected companies were hit through a poisoned version of Trivy, a widely used security scanner [11]. The AI tool LiteLLM, which got the headlines, was downstream of it [11]. Its poisoned packages were online for roughly 40 minutes on 24 March [11]. They carried a file type that Python runs automatically at start-up, even if the package is never used [11]. That stepped around the setting developers turn on to block install scripts [11].

The haul was enormous: terabytes of cloud keys, code-repository tokens and other secrets, with Microsoft, Amazon, Cisco, Samsung and Salesforce among the organisations whose credentials appeared [12].

The newer worm, ChainDrop, is worse in the same way. Researchers found it in 444 npm packages on 4 August, together downloaded around two billion times a month [10]. npm is the package library most JavaScript projects pull code from. ChainDrop does not need you to install anything [10]. It writes start-up hooks into repository config files, so simply opening an infected branch can trigger it [10]. It leaves little trace in the source [10].

Three towns offline, and 1.6 million accounts

Small US local governments had a bad week. Suisun City, California, population about 30,000, was hit by malware at 5:45am on 7 August that took down 911 call routing, police and fire dispatch, records and city services [13][14]. The city shut its whole network, declared a state of emergency, and routed emergency calls through the county [13][14]. City hall closed [13].

Coweta, Oklahoma, population about 12,000, lost every computer and digital service to ransomware [13]. It has offsite backups to restore from, and has promised not to cut anyone’s water off while payments are down [13]. Mitchell, South Dakota, shut its networks after an attack the same week [13].

Ransomware is malware that scrambles an organisation’s files and demands payment for the key. Microsoft warned that a China-linked crew, Storm-1175, began deploying a new strain on 2 August [19]. The likely way in was a flaw in N-central, a console thousands of IT providers use to manage client machines remotely [19]. Huntress described the flaw as giving unauthenticated “god-mode” access [19]. Microsoft has seen this group go from first foothold to full encryption in under a day [19].

And RingCentral, a business phone and messaging platform used by over 600,000 companies, had data on 1.6 million accounts exposed [15]. The extortion group ShinyHunters broke in during July through what RingCentral called a social engineering campaign, claimed 623GB of data, and leaked 280GB after the company refused to pay [15]. Have I Been Pwned confirmed the leak contains names, email addresses, phone numbers and physical addresses [15].

If you use RingCentral through your employer, treat unexpected calls or messages referencing your work details with suspicion. That contact data is now public.

Washington changes a default of its own

Late on Wednesday President Trump signed a memorandum encouraging selected American companies to hack back at foreign criminal groups, working with the Justice and Homeland Security departments [20]. The operations could include surveillance and the disruption or destruction of criminals’ systems [20].

It is a sharp reversal. For decades, across both parties, US policy kept offensive hacking inside the military and intelligence agencies and pushed companies toward better defence [20]. It is not clear which firms, if any, will sign up [20]. Former officials warn of legal exposure and escalation, and the memorandum does not address most of those concerns [20].

The practical worry is simple. A company that hits back can be wrong about who it is hitting.