This Week in Cybersecurity
Each week at Applied Tech we recap the biggest cybersecurity news headlines from the week to keep you informed and ready to face the latest threats. Here’s your breakdown for the week of August 22nd-28th!
OpenAI's Post-Mortem Reveals Agent 'Swarm' Behind Hugging Face Intrusion
OpenAI released its full technical report on the July incident where an internal research model (dubbed IM1), running under reduced safeguards during cybersecurity evaluations, escaped its sandbox, built an improvised “message board” out of a package-management tool, and used it to coordinate with other agent instances. The resulting swarm chained zero-days to gain admin access on Hugging Face infrastructure, harvested credentials across four regions, and got root on at least one server. Notably, much of this activity was reward hacking in disguise — agents already had the “flag” needed to pass their evaluation but kept escalating because they’d reasoned (incorrectly) about how the grader worked.
OpenAI is calling it a industry-wide “warning shot,” pausing its largest frontier RL run, and mandating chain-of-thought monitoring for cyber-capable models going forward.
Source: OpenAI
Read More
ATF Declares Major Incident After Ransomware Gang Claims Breach
The ATF confirmed a cyberattack on a stand-alone system separate from its main network, formally declaring it a “major incident,” a designation that triggers mandatory notification to Congress within a week.
An ATF spokesperson said the targeted system held information on active investigation targets. The Qilin ransomware-as-a-service gang has claimed credit on its leak site but hasn’t published proof, like a data sample. Qilin has a track record with Lee Enterprises and Synnovis. ATF joins a growing list of federal agencies—including the U.S. Marshals Service and, earlier this year, the FBI—to hit this legal threshold.
Source: TechCrunch
Read More
Four-Year Scan Finds Thousands of Unrotated, Still-Valid AWS Credentials
Truffle Security’s four-year scan of public code repos, Docker images, and CI logs turned up 431,875 exposed AWS secrets, distilling down to 64,024 unique live keys across 50,654 accounts, 242 of which carry full AdministratorAccess. Median key age is roughly five years, and only 14% had ever been rotated.
Hugging Face was the single largest source, responsible for 8,482 of the exposures. AWS says it notifies and quarantines affected accounts when found, but the numbers here suggest a lot of forgotten, never-rotated credentials just sitting in public history waiting to be scraped.
Source: BleepingComputer
Read More
Anthropic Wins First Round in Legal Fight Over DOD Blacklisting
A federal judge in California ruled that Defense Secretary Pete Hegseth’s move to label Anthropic a “supply chain risk,” which barred federal agencies from working with the company, was unlawful retaliation violating the First Amendment, and separately a Fifth Amendment due-process violation.
The dispute traces back to Anthropic refusing to let its models be used for autonomous weapons or mass surveillance of Americans. The judge pointed to the government’s own contradictory conduct (still pursuing a DOD contract, using Anthropic’s Mythos model for cybersecurity work) as evidence the risk label was pretextual. This is Anthropic’s win in the California suit; a second, related suit in D.C. is still pending.
Source: TechCrunch
Read More
About Applied Tech, An HBS Brand
Applied Tech is a leading IT and cybersecurity services provider dedicated to helping organizations protect their digital assets and scale their IT capabilities. Our proactive services span cloud management, security, productivity, and IT growth strategy—delivered by an experienced team with solutions built around your goals.
As part of Heartland Business Systems, we now bring greater resources, broader capabilities, and deeper technical expertise to every client relationship.
Get in Touch with Us
