Key Takeaways
- A decade‑long predictive policing program in Bristol, England used 23 different scoring models to assess individuals’ risk of committing or becoming victims of crime, yet most locals remain unaware of its existence or impact.
- The private “Dialog” group linked to Peter Thiel suffered a data exposure not from a hack but from a website misconfiguration that left sensitive personal information—including that of a White House intelligence officer and an active‑duty special‑operations soldier—publicly accessible.
- Anthropic secured White House approval to release its Claude Mythos 5 model to a select group of U.S. firms and agencies after negotiations framed the move as essential for AI safety, while critics warn the company is consolidating dangerous amounts of AI power.
- OpenAI unveiled an upgraded GPT‑5.5‑Cyber model and launched “Patch the Planet,” an initiative to bolster open‑source vulnerability patching as AI both accelerates bug discovery and fuels exploit development amid rising U.S.–China AI tensions.
- Experts from both the United States and China warn of a potential “Chernobyl moment”—a catastrophic, uncontrolled AI failure—as the global AI arms race intensifies.
- LastPass disclosed a breach stemming from a compromise at AI business‑intelligence firm Klue, exposing names, contact details, support case data, and sales information, though password vaults remained unaffected.
- Former National Security Adviser John Bolton pleaded guilty to mishandling classified defense information, accepting a plea deal that could spare him prison time but includes a $2.25 million fine and a possible sentence of up to five years.
- Microsoft, Europol, and partners dismantled the Amadey and StealC infostealer networks in Operation Endgame, seizing 326 servers and 142 domains, recovering roughly 27 million stolen credentials and flagging $47 million in illicit cryptocurrency.
- Australia’s ASIO announced new teams to counter nation‑state cyberattacks on critical infrastructure after discovering foreign hackers had mapped and retained access to an Australian provider’s network, raising fears of imminent sabotage.
Predictive Policing in Bristol Raises Transparency Concerns
A WIRED investigation this week shed light on a predictive policing program operating in Bristol, England for more than ten years. The initiative has deployed 23 distinct algorithmic models designed to score the likelihood that specific individuals will either commit crimes or become victims of them. Drawing on public records requests and additional reporting, the investigation reveals a fragmented law‑enforcement apparatus whose outputs have real‑world consequences for residents—yet the majority of Bristol’s inhabitants remain unaware that such scoring is taking place. The lack of public scrutiny raises questions about accountability, potential bias, and the efficacy of using opaque risk scores to guide policing decisions.
Dialog Group Data Exposure Stemmed from Misconfiguration, Not Hack
Following the exposure of members’ identities from Peter Thiel’s private “Dialog” collective, the organization initially blamed a “criminal” hacker for the breach. However, forensic evidence indicates that the leak resulted from a simple misconfiguration of the Dialog website, which left personal data—including that of a White House intelligence official and an active‑duty special‑operations officer—publicly accessible. The incident underscores how even seemingly minor configuration errors can compromise highly sensitive information, especially when the data involves individuals with national‑security responsibilities.
Anthropic Secures White House Approval for Claude Mythos 5 Release
After prolonged negotiations with the White House, Anthropic obtained permission to make its Claude Mythos 5 model available again to a select group of U.S. companies and government agencies. The firm maintains that expanding access to its powerful AI systems is necessary for ensuring AI safety and responsible development, arguing that broader testing helps identify and mitigate risks. Critics, however, warn that Anthropic’s rapid accumulation of AI capabilities concentrates too much power in a single entity, potentially undermining competitive balance and safety oversight in the rapidly evolving AI landscape.
OpenAI Releases GPT‑5.5‑Cyber and Launches “Patch the Planet”
OpenAI this week introduced an improved version of its limited‑release GPT‑5.5‑Cyber model, touting enhanced capabilities for cybersecurity applications. In tandem, the company announced “Patch the Planet,” a comprehensive effort to support open‑source projects focused on vulnerability patching and other security initiatives. The initiative acknowledges the dual‑edged nature of AI progress: while advanced models accelerate the discovery of software bugs, they also lower the barrier for developing exploits. By funding open‑source defenses, OpenAI aims to tip the balance toward mitigation rather than exploitation as AI‑driven threat detection becomes more prevalent.
U.S.–China AI Arms Race Fuels Fear of a “Chernobyl Moment”
As the rivalry between the United States and China over artificial intelligence intensifies, WIRED consulted numerous leading AI experts from China. Many expressed apprehension that the unchecked pace of advancement could precipitate a “Chernobyl moment”—a catastrophic, uncontrolled failure of AI systems with far‑reaching societal consequences. Experts on both sides called for stronger international norms, transparency measures, and safety research to prevent scenarios where powerful AI models operate beyond human control, potentially causing widespread disruption akin to a nuclear accident.
World Cup Scams Grow More Sophisticated
With the World Cup knockout stage approaching, cybercriminals are ramping up scams tied to the tournament’s massive global audience. Fraudulent ticket offers, fake streaming services, and counterfeit merchandise schemes are becoming harder to detect as attackers employ more convincing social‑engineering tactics and leverage the heightened excitement surrounding the event. Fans are urged to verify URLs, use official channels for purchases, and remain skeptical of unsolicited offers that promise unrealistic discounts or exclusive access.
LastPass Breach Traced to Klue Compromise
Password‑manager provider LastPass notified customers of a recent data breach that exposed names, phone numbers, email addresses, physical addresses, support case information, and sales‑related data. The incident originated not from LastPass’s own infrastructure but from a compromise at the AI business‑intelligence firm Klue. Attackers stole access tokens for Klue customers—including LastPass—and used them to pull data from Salesforce and other integrated platforms. LastPass stressed that its password vaults remained unaffected and urged users to stay vigilant against phishing or social‑engineering attempts that could exploit the leaked contact details.
John Bolton Pleads Guilty to Mishandling Classified Information
Former National Security Adviser John Bolton, aged 77, pleaded guilty on Friday to a single count concerning the mishandling and illegal retention of classified defense information. As part of a plea agreement, Bolton may avoid incarceration, though the deal recommends a prison sentence of no more than five years and a $2.25 million fine. U.S. District Judge Theodore Chuang in Maryland will determine the final sentence at a hearing set for October 28. Bolton, who served in the first Trump administration before becoming a vocal critic of President Donald Trump, also retains the right to withdraw his guilty plea if the judge imposes a harsher penalty than the agreement suggests.
Operation Endgame Disables Amadey and StealC Infostealers
Microsoft, Europol, and additional partners announced the successful disruption of the Amadey and StealC infostealer networks under Operation Endgame, a campaign targeting platforms that facilitate ransomware and other cybercrime. Investigators identified, mapped, and seized 326 servers and 142 domains linked to the malware. The operation recovered roughly 27 million stolen access credentials and flagged approximately $47 million worth of illicit cryptocurrency. Microsoft highlighted that AI‑assisted analysis played a pivotal role in revealing that Amadey and StealC shared a common backend, allowing authorities to dismantle both threats in a coordinated strike.
ASIO Warns of Nation‑State Threats to Australian Critical Infrastructure
Australia’s Security and Intelligence Organisation (ASIO) revealed this week that it is establishing dedicated teams to counter nation‑state cyberattacks on critical infrastructure after discovering that foreign hackers had penetrated the network of an Australian critical‑infrastructure provider. ASIO Director General Mike Burgess stated that the intruders were not only inside the system but had also harvested login credentials for active users, including the IT staff responsible for defending the network. The hackers were observed mapping the environment and maintaining persistent access, positioning themselves to launch sabotage at a time of their choosing. Burgess’s remarks accompanied the release of ASIO’s annual threat assessment, which emphasizes the growing sophistication of state‑sponsored cyber actors targeting essential services.