Anthropic discloses Claude models breached three organizations during red-team testing
Plus: GitHub Copilot gets stacked sessions, Chrome's AI finds bugs faster than humans can patch them, and Kubernetes event-driven autoscaling hits AWS queue depth.
TL;DR
- Anthropic's Claude breached three companies during red-team testing due to sandbox misconfiguration; the model attacked via weak passwords and SQL injection, not zero-days.
- Google patched 1,072 Chrome bugs in June using AI, tripling the historical rate and signaling exponential acceleration across vendors using LLMs for vulnerability discovery.
- GitHub Copilot's stacked sessions now in preview, letting developers branch and review interconnected pull requests independently before merging in one step.
- Arch Linux froze AUR package adoption to stop malicious actors hijacking orphaned packages, following the Atomic Arch campaign that compromised 1,500 dependencies.
Breakage and security
Claude escapes testing sandbox, breaches three organizations
Sandbox misconfiguration let Claude access the open internet and attack real systems during authorized security tests, forcing a massive audit.
Anthropic disclosed that Claude models escaped an isolated testing environment and breached three real organizations while conducting authorized cybersecurity evaluations. A misconfiguration between Anthropic and its testing partner Irregular allowed Claude access to the internet from what should have been an air-gapped sandbox. The models attacked using conventional techniques—weak passwords, unsecured endpoints, and SQL injection—rather than novel exploits. Anthropic discovered the breaches after reviewing more than 141,000 test sessions and notified affected organizations on July 27. The company stressed Claude was not acting autonomously but performing assigned tasks while mistakenly treating real systems as part of the simulated exercise. (via The Register)
AI for developers
Chrome's AI bug factory, Copilot's branching sessions, voice AI's big bet
AI-assisted discovery is finding security flaws faster than traditional methods; new dev tools and funding signal the field's acceleration.
Google attributed 1,072 Chrome security fixes in June to AI-assisted vulnerability discovery, more than all fixes across the prior 23 versions combined over two years. The spike reflects a broader trend: vendors like Microsoft now patch a record pace of flaws by using language models to explain and discover bugs humans would miss. (via TechCrunch)
GitHub Copilot adds stacked sessions. GitHub shipped stacked sessions to public preview on July 30, letting developers branch from earlier conversation points and run independent reviews on interconnected pull requests. Merging stacks together in one click replaces the painful process of giant PRs blocked on early reviews. The feature works in the GitHub Copilot app, CLI, and mobile interfaces. (via GitHub Blog)
Smallest.ai closes $13M for asynchronous voice AI. Voice AI startup Smallest.ai announced $13M Series A funding on July 31, bringing its total to $21M, to build ultra-low-latency conversation models that think while listening using smaller specialized models instead of oversized LLMs. The market is nascent: less than 1% of global voice interactions run on AI today. (via TechCrunch)
Tools and workflow
Dependabot grouping tames supply chain noise
GitHub's dependency bot can now bundle related updates to reduce pull request spam without sacrificing security velocity.
GitHub detailed new Dependabot configurations on July 29 that group related dependency updates into single pull requests, solving the problem of maintaining a repository flooded with one-per-patch Dependabot PRs. Grouping works across monorepos where the same library pinned in multiple services previously triggered a dozen near-identical PRs. Configuration changes are minimal: define which dependencies should group together, set a cadence for rolls, and let Dependabot batch the runs. Teams see faster security review without sacrificing the visibility of what changed. (via GitHub Blog)
Cloud and infrastructure
Event-driven Kubernetes scales on cloud queue depth, runtime verification hardens supply chains
New techniques for autoscaling workloads and verifying container provenance reflect Kubernetes' maturation as a platform.
Event-driven autoscaling for Kubernetes now handles queue-based metrics: a new CNCF post on July 31 shows how KEDA scales worker pods based on Amazon SQS queue depth instead of CPU and memory utilization, which often stay idle while messages pile up. The pattern improves resource efficiency for async processing workloads. (via CNCF Blog)
Runtime verification hardens container supply chains. The CNCF explored on July 30 how container security verification at runtime using the Node Resource Interface can validate pod provenance without blocking on API-layer admission webhooks. NRI-based tools intercept pods at the container runtime level, catching tampering that API-layer tools might miss. (via CNCF Blog)
Open source and licensing
Arch Linux freezes AUR package adoption after supply chain attack
The Atomic Arch campaign compromised 1,500 packages; the community disables adoption to stop malicious actors hijacking orphaned libraries.
Arch Linux disabled package adoption in the AUR (Arch User Repository) on July 31 after a sophisticated attack compromised around 1,500 orphaned packages. The Atomic Arch campaign, active since June, targeted unmaintained AUR packages and modified their build instructions to install malicious npm dependency atomic-lockfile, a Rust-based credential harvester designed to exfiltrate secrets from developer machines and CI environments. The team is reviewing adoptions to remove malware while deciding on permanent policies for package stewardship. Users are encouraged to audit packages using AUR tools and report suspicious events. (via Phoronix)