This article is more than 1 year old

US officials, experts fear China ransacked Exchange servers for data to train AI systems

Plus: T-Mobile US apologizes, security holes found in medical pumps, and more

In brief The massive attack on Microsoft Exchange servers in March may have been China harvesting information to train AI systems, according to US government officials and computer-security experts who talked to NPR.

The plundering of these Exchange systems was attributed to Chinese government cyber-spies known as Hafnium; Beijing denied any involvement.

It's said the crew exploited four zero-days in Redmond's mail software in a chain to hijack the servers and siphon off data. And what started small turned into what Chang Kawaguchi, CISO for Microsoft 365, told NPR this month was the fastest scale-up of a cyber-attack he'd ever seen.

US government officials, and those in the infosec industry, are apparently concerned that, given the wide range of organizations targeted – from big biz to shops, dentists, and schools – the Chinese government could be trying to train machine-learning systems on mountains of Americans' messages, calendars, and files.

And this Exchange harvesting is on top of the huge databases of personal information already swiped from the US government and the private sector.

"The Chinese have more data than we have on ourselves," William Evanina, a former director of the National Counterintelligence and Security Center, was quoted as saying.

"So you have the OPM data breach," he continued, "you have an entire security clearance file for someone, you have Anthem records, you have his Marriott point record, credit cards, Equifax, his loans, his mortgages, his credit score. They know everything about you before they even bump you on a cruise or on a vacation."

Evanina spoke more on the threat from China here [PDF] before the Senate intelligence committee at the start of August, if you're interested.

We hope you've patched ProxyToken, aka CVE-2021-33766, in July's Patch Tuesday patch from Microsoft for Exchange server, because details on how to exploit it to steal people's emails are now public.

Homeland Security gets smarter about recruiting security talent

While the media's attention was focused last week on President Biden meeting tech CEOs to talk about computer security, Homeland Security was focusing on the essentials: getting the right staff.

The department is asking for comment on its proposed Cybersecurity Talent Management System (CTMS), which will revamp how the agency hires infosec bods by allowing for people with unconventional qualifications and backgrounds to apply and adjusting government pay scales to ensure folks get "sufficiently competitive, market-sensitive compensation."

These are two massive sticking points for those wishing to sign up for government service. It's not that long ago that if you wanted to get onto the FBI's cyber-security team you had to undergo the physical training course aimed at agents, and anyone who'd even once taken illegal recreational substances was instabanned from joining, though the Feds eased up just a little on that one.

On the pay front, security engineers earn big bucks in the private sector. It takes a very committed citizen to sign up for government service when they could get double the salary and stock options in Big Tech. Obviously the government isn't going to be able to match FAAMG wages, but every little bit helps.

Speaking of Uncle Sam... The White House on Monday launched the US Digital Corps, a two-year fellowship program that finds junior software engineers, data scientists, and other geeks roles at federal agencies.

Palantir and that 'glitch'

After some FBI agents were able to access case paperwork that should have been off limits in a Palantir database, the software maker has said it was due to the Feds not RTFM'ing as opposed to a glitch that was reported in the media.

"There was no glitch in the software," Palantir told The Register. "Our platform has robust access and security controls. The customer also has rigorous protocols established to protect search warrant returns, which, in this case, the end user did not follow."

What happened is this: the FBI was investigating Virgil Griffith, who is accused of helping North Korea evade sanctions through the use of crypto-coins. Last March, the FBI got a search warrant to scour his Facebook and Twitter accounts for info and uploaded it to the Feds' Palantir database for analysis. However, it appears the agents forgot something: they allowed any other agent who could access the Palantir system to view Griffith's records, rather than securing the data to select, relevant personnel only.

"The government understands that this outside access to the search warrant returns was made possible because, when data is loaded onto the platform, the default setting is to permit access to the data to other FBI personnel otherwise authorized to access the platform," wrote [PDF] Audrey Strauss, US Attorney for the Southern District of New York.

"When the search warrant returns here were loaded onto the platform, those default settings were not changed to restrict access to the search warrant returns to the FBI personnel actually engaged in reviewing the search warrant returns pursuant to the warrants."

Earlier this month an FBI agent unconnected with the Griffith case mentioned he'd seen the data. Further investigation showed three FBI analysts and a now-retired agent accessed the social media information when they shouldn't have. The Feds are now checking if this permissions issue is more widespread.

If you remember the Razer driver issue, it turns out there may be as many as 2,500 USB device models that potentially auto-run an interactive installer with high privileges when a gadget is inserted into a Windows PC. Those installers could be used by a local logged-in user to gain SYSTEM-level powers. The problem is that Windows allows these interactive installers to run automatically regardless of the scope of access the present user has.

T-Mobile US CEO sheds a little more light on database raid

Earlier this month T-Mobile US admitted the personal information of 48 million of its subscribers had been stolen, and now CEO Mike Sievert has partially explained how it was robbed.

"In simplest terms, the bad actor leveraged their knowledge of technical systems, along with specialized tools and capabilities, to gain access to our testing environments and then used brute force attacks and other methods to make their way into other IT servers that included customer data," he said. "In short, this individual’s intent was to break in and steal data, and they succeeded."

He added that he was "truly sorry." The telco will offer the usual two years of free identity protection services to customers, among other things. To try and stop this happening again, T-Mobile US is hiring infosec outfit Mandiant and consultants from KPMG to harden up its systems against attack. A person claiming they were behind the data theft told the Wall Street Journal the cellular network giant's security was "awful."

Heart-stopping security flaws

For once that's not an exaggeration: McAfee uncovered five serious flaws in medication infusion pumps made by Germany's B. Braun.

Infusion pumps have been in use for decades, administering set dosages of medicine to a patient without needing a medic on hand, and they are increasingly network controlled. While exploitation was non-trivial, the teams from McAfee and security shop Culinda demonstrated it would be possible to adjust the dosage levels the pump would provide remotely it it was connected to a hospital network.

"The ability to remotely manipulate medical equipment undetected, with potential for patient harm, is effectively weaponizing these point of care devices," said Shaun Nordeck, an interventional radiology resident physician at a level-one trauma center.

"This is a scenario previously only plausible in Hollywood, yet now confirmed to be a real attack vector on a critical piece of equipment we use daily." ®

More about

More about

More about


Send us news

Other stories you might like