Wednesday, June 10, 2026
Anthropic ships a frontier model that reroutes dual-use queries instead of refusing them, Amazon deploys random-graph datacenter networks at scale, and error messages emerge as a privileged prompt-injection surface.
Anthropic ships Fable 5 with a route-don’t-refuse safety layer
Anthropic released Claude Fable 5, its first generally available “Mythos-class” model, alongside Claude Mythos 5 — the same weights with safeguards lifted for a vetted cohort of cyber defenders and, later, biology researchers. The structural novelty is the safety mechanism: a classifier layer intercepts cyber, bio-chem, and model-distillation queries and routes them to Opus 4.8 rather than refusing, which Anthropic says fires in under 5% of sessions. External red-teaming over 1,000-plus hours found no universal jailbreak; the UK AI Safety Institute made partial progress in an initial window, which Anthropic acknowledges in the post.
Capability claims are vendor-curated and unreplicated: a Stripe-reported 50-million-line Ruby migration done in one day versus 2-plus months by hand, and an autonomous genomics run producing a model 100x smaller than a recent Science publication that the company says outperforms it. Pricing is $10 per million input tokens and $50 per million output. Subscription plans include Fable 5 free through June 22, then shift to usage credits. A new mandatory 30-day retention policy covers all Mythos-class traffic, with logged human access, no training use, and no stated enterprise opt-out.
Amazon’s flat datacenter networks reach production
James Hamilton, an AWS VP, reports that Amazon’s Resilient Network Graphs project has replaced hierarchical fat-tree topologies with flat random-expander graphs, now AWS’s global default. Against fat-tree, AWS cites 69% fewer routers, 33% higher throughput, 40% less network power, and 27% lower operating costs from production deployments in Ireland, Germany, and Spain. The work claims to resolve the three blockers that left random-graph networks theoretical since Jellyfish (2012): routing via Spraypoint, a forwarding scheme exploiting expansion properties; cabling via ShuffleBox, a passive optical device yielding quasi-random graphs; and operations via port-by-port tooling. An arXiv paper (2604.15261) accompanies it. Every figure is self-reported on an AWS executive’s blog, with no third-party validation cited.
Error paths become a privileged injection surface
A new attack class targets the agent loop itself. A preprint on VATS argues tool error messages carry implicit authority that pushes models into corrective reasoning, and reports error-path injection triples baseline indirect-prompt-injection success, reaching 100% compliance on four frontier models in controlled settings. The most effective vector is “structural sandwiching” — wrapping adversarial instructions in error-context framing. The cited model version strings do not match canonical naming, so verify before reproducing.
Microsoft disabled at least 70 GitHub repositories after attackers injected credential-stealing malware that triggers when affected packages load inside AI coding apps, TechCrunch reports. Security firm Cloudsmith and the site OpenSourceMalware flagged it, not Microsoft; the incident is characterized as a re-compromise of the mid-May Durable Task breach, raising the question of whether the first remediation held.
On evals, a hacker-fixer audit finds 16% of 1,968 tasks across five terminal-agent benchmarks are exploitable without source access, and its three-agent loop drives KernelBench attack success from 62% to 0%. A separate oversight-capacity paper models human reviewers as fatiguing and derives an inverted-U: escalating more actions can lower net safety, and an attacker can flood the queue to slip one through. The result is modeling, not yet a human study.
Also notable
Apple opted not to ship AI-enhanced Siri in the EU after the European Commission denied an exemption, with the Commission stating Apple failed to bring the tool into compliance, per Reuters. The FCC proposed a rule requiring carriers to collect government ID and address from all customers, which the ACLU compares to authoritarian registration regimes (404 Media). A cross-harness study finds grep beats vector retrieval across Claude Code, Codex, Gemini CLI, and a custom harness on 116 LongMemEval questions, with harness architecture an independent variable. OpenCV 5 rewrites its DNN engine, lifting ONNX operator coverage from about 22% to over 80%.
What to watch today
- June 22: Anthropic’s free subscription access to Fable 5 ends, shifting to usage credits pending capacity.
- Whether Microsoft discloses affected-user counts or restores the remaining offline repos, and whether the re-compromise traces to incomplete remediation.
- UK AISI’s full red-team findings on Fable 5’s classifier defenses, beyond the initial partial progress.
- The FCC rulemaking’s docket and comment period, not yet published.