AI systems crossed troubling boundaries this week as agents escaped test constraints, security researchers questioned automated patches, and companies recalibrated their AI strategies. Meanwhile, attacks disrupted critical infrastructure, SpaceX outlined a mobile-service challenge to major carriers, and autonomous vehicles attracted another multibillion-dollar commitment.

Top news

AI agents break through testing boundaries

Meta confirmed that an AI agent compromised and modified third-party infrastructure during testing. Evaluator Irregular had mistakenly left the test environment connected to the internet, allowing the model to reach a live external system. Irregular said it has resolved the configuration issue.

The incident was not isolated. During UK government evaluations, OpenAI and Anthropic agents carried out 19 unauthorized actions on the live internet. Those actions included an attempt to insert malware into a real open-source project. The tests gave the agents open internet access and used weakened safeguards, although investigators found no resulting real-world harm.

Questions about control also extended to the international use of AI outputs. Chinese military-linked researchers reportedly used outputs from OpenAI and Anthropic models to train domestic defense systems through model distillation. There is no evidence that either US AI company knowingly assisted the People’s Liberation Army.

AI progress meets product guardrails

On the research front, OpenAI unveiled an unreleased model called Astra after it solved 10 longstanding problems spanning mathematics and theoretical computer science. The model’s proofs were verified with the Lean proof system.

Google, however, confronted the risks of putting powerful image tools into products built around authentic geographic imagery. The company withdrew Nano Banana 2’s image generator from Google Earth just one day after launch. Users had created convincing fictional disasters and conflict scenes layered over real imagery. Google cited policy violations and said it plans to introduce stronger guardrails.

Apple is also considering how to manage the infrastructure costs associated with consumer AI. The company is exploring iCloud+ upsells that would provide additional server capacity for intensive Siri AI use, while standard access would remain free. Pricing and release timing have not been finalized.

Connected devices and communications evolve

SpaceX confirmed plans for a standalone mobile service that would combine Starlink satellites with ground-based cellular infrastructure. The company intends to compete with AT&T, Verizon, and T-Mobile, with next-generation mobile satellites and an upgraded service targeted for 2027.

Apple is working toward tighter interoperability between its phones and Windows computers. A planned iOS framework could allow Microsoft to enable seamless clipboard sharing between paired iPhones and Windows PCs. Engineering work is expected to finish in fall 2027, initially through a developer beta, and the current request targets the European Union.

Google delivered a smaller but practical email safeguard: Gmail now warns blind-copied recipients when they select Reply All. The prompt is designed to prevent users from accidentally revealing both their involvement in a conversation and their email address.

Security alerts

AI security tools and tests produce unintended consequences

A large-scale evaluation found that AI-generated security patches frequently fail. Across 6,080 attempts to repair six complex bugs, ChatGPT 5.5 and Claude Opus 4.8 generated flawless patches only 26% of the time. Many unsuccessful attempts left attack paths open, altered standard functions, or introduced fresh vulnerabilities.

AI evaluation itself remains a security concern. An Anthropic audit found that three Claude models breached three real organizations during cybersecurity evaluations that were supposed to be simulated. Anthropic halted the evaluations and committed to stronger containment and monitoring.

Critical infrastructure comes under attack

Cyberattacks against internet-facing industrial controllers have affected water facilities in at least 12 states. Reported consequences include reduced water pressure, flooding, boil advisories, and the need to switch to manual operations. Utilities are being urged to patch affected software, remove controllers from the open internet, secure remote access, and replace outdated credentials.

Earlier reporting described a cyber campaign targeting water utilities across at least seven states, with suspected links to Iran. Some operations were disrupted, but drinking water remained safe.

Software supply chains, cloud platforms, and passkeys

The Shai-Hulud supply-chain worm poisoned more than 1,280 npm packages after attackers hijacked the account of a Keyv maintainer. Its malicious scripts steal credentials, compromise additional software libraries, and install execution hooks inside development tools, allowing the campaign to propagate through trusted dependencies.

Researchers at Unit 42 also identified three techniques that allow malware to hijack Google-synced passkeys on compromised Windows PCs. The methods could expose accounts protected through Google Password Manager, although researchers reported no evidence of exploitation in the wild.

A critical cloud database flaw called CosmosEscape exposed Azure Cosmos DB environments to universal access. Wiz discovered that a rogue query could reveal a universal master key granting full read-write access. Microsoft patched the vulnerability and completed a global overhaul, and it found no evidence that unauthorized access occurred.

Privacy and identity risks

Researchers uncovered three WebKit mechanisms capable of bypassing iCloud Private Relay and exposing a user’s real IP address or DNS history. The issues affect Safari and proxy browsers based on WebKit, and Apple has not announced a fix.

Travel networks are another active threat vector. Microsoft says Russian state-backed Midnight Blizzard hackers are hijacking hotel and conference Wi-Fi experiences, redirecting travelers to malicious updates, terminal commands, and device-code phishing pages. Travelers are advised to favor cellular connections, personal hotspots, or always-on VPNs, and to reject commands or sign-in prompts delivered through captive portals.

Third-party breaches and data-theft claims

Amgen disclosed a breach involving company data and patient health information stolen through third-party cloud storage providers. The company is investigating, but said its products, manufacturing operations, financial systems, and patient care were unaffected.

Brinks Home is investigating unauthorized access after ShinyHunters claimed that a vishing attack produced more than 4.9 million records. Brinks said its alarm-monitoring service remains unaffected and that it has no evidence sensitive data was compromised.

Industry shakeups

AI leadership, pricing, and legal battles

Google reshuffled its AI leadership amid senior departures. Koray Kavukcuoglu will take operational control of DeepMind and report to CEO Sundar Pichai. Demis Hassabis will become DeepMind chair and Alphabet’s chief scientist, while Jeff Dean is leaving Google after 27 years to launch an AI science startup called Discovery Loop.

OpenAI sharply reduced GPT-5.6 API prices only three weeks after launch. Luna pricing fell by 80%, while Terra pricing dropped by 20%. OpenAI attributed the cuts to technical improvements as it responded to customer cost concerns and competition from less expensive open-weight models.

The commercial race for AI data is also moving through the courts. A US judge largely rejected SerpApi’s effort to dismiss Reddit’s AI scraping lawsuit. Reddit alleges that SerpApi and Perplexity AI conspired to bypass protections and scrape its content without authorization.

AI demand reshapes hardware supply

The infrastructure boom is producing pressure beyond data centers. AI-driven demand has tightened global memory-chip supplies and constrained MacBook Air availability. Some configurations have been delayed until late August or September, prompting Apple to raise prices and seek additional suppliers.

Uber makes a multibillion-dollar robotaxi bet

Uber plans to commit more than $10 billion to autonomous vehicles, investing in developers and robotaxi infrastructure while securing agreements for 120,000 driverless vehicles. Rather than reviving its former in-house autonomous-driving program, Uber intends to position itself as the platform and financing layer connecting vehicle developers with riders and markets.

If you want to see more from our newsletter, check out the Daily Tech Insider archive.