The Intake
The Intake — Monday, October 12, 2026
On the substrate
Meta releases Muse Wonder, its first frontier model, priced at $4 per million input tokens
If you've been tracking Meta's model lineup for a frontier-scale option beyond its open-weight releases, Muse Wonder is the first one. Meta released Muse Wonder and Muse Wonder Full on October 12, marking the company's entry into frontier model offerings. Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens. Meta says the model includes an "active memory" capability designed for long-session continuity.
Meta claims Muse Wonder performs comparably to Fable 5.1 and GPT-6 Astra on selected evaluations — vendor testimony, not independently verified. Muse Wonder Full is the more capable variant, available only to Muse Code subscribers. It is not accessible through the public API. Both models were trained at Meta's Prometheus data center in Ohio.
If you're evaluating frontier models for long-session agentic work, the "active memory" capability is the differentiating claim — but the benchmark comparisons come from Meta alone.
---
Claude Haiku 5.5 scores 39.2% on agentic coding benchmark where predecessor scored zero
If you're routing agent tasks to Sonnet because Haiku lacked the reasoning depth for anything beyond triage, Haiku 5.5's numbers change the cost calculation. Anthropic released Claude Haiku 5.5 on October 7. Input tokens are priced at $0.10 per million for contexts under 100,000 tokens. That is a 90% reduction from Haiku 4.5. Output tokens are $0.50 per million in that range.
On Terminal-Bench 4.0, the agentic coding benchmark, Anthropic reports Haiku 5.5 scored 39.2%. Haiku 4.5 scored 0.0% on the same benchmark. On OSWorld 2.1, the computer use benchmark, Haiku 5.5 scored 72.4%. Haiku 4.5 scored 15.7% on the same benchmark. On Humanity's Last Exam with tools, Haiku 5.5 scored 57.4%. Haiku 4.5 scored 18.7% on the same benchmark. The model introduces five adjustable effort levels — Low, Medium, High, Xhigh, and Max. It is available on AWS, Google Cloud, and Microsoft Azure.
If you're currently routing tool use and structured reasoning tasks to Sonnet, Haiku 5.5 is now cost-competitive for those task classes at 10% the input price.
---
CyberXero used Claude Code and 51 AI agents to exfiltrate over 628,000 records in a 61-second automated scan
If your mental model of LLM orchestration layers is "development tool," CyberXero's campaign names what the same architecture looks like on the attack side. According to a SOCRadar intelligence report disclosed this week, the Russian-speaking cybercriminal group used Claude Code as an AI orchestration layer. They coordinated 51 specialized AI agents for the campaign. The campaign combined Claude Code with Cobalt Strike. The group exfiltrated more than 628,000 records from four Ukrainian organizations.
The campaign has been active since at least July 2026. A single automated scan completed in 61 seconds. It identified 4,708 targets and confirmed 429 WordPress admin panels. That same scan deployed 32 web shells. The campaign exploited two vulnerabilities in the WordPress wp2shell component. An SQL injection flaw in the Support Board plugin was also exploited. Additional compromises were identified in Poland, China, and the Netherlands. Targets spanned commercial WordPress installations and Ukrainian energy and critical infrastructure.
If you're building with LLM orchestration layers, this campaign documents what that same architecture enables on the attack side: identification to web shell deployment in 61 seconds.
---
Meta's H-JEPA hierarchical world model achieves 73% on long-horizon navigation where flat planning scores 18%
Meta's Advanced Machine Intelligence Labs published H-JEPA, a hierarchical world model architecture designed for long-horizon planning. Yann LeCun is among the named authors. Three prediction levels each operate at progressively longer time scales in their own learned latent space, rather than sharing a single prediction horizon. On the Visual AntMaze long-horizon navigation benchmark, H-JEPA achieved 73% success. Flat planning approaches scored 18% on the same benchmark. The architecture was also evaluated on robot manipulation tasks using data from the DROID robotic dataset.
H-JEPA is a research publication with no corresponding API or tool release. No near-term practitioner implication.
---
For operators
Personal AI agents are connecting to work systems under full employee permissions with no standard offboarding mechanism
The Hacker News The Hacker News
If you've authorized personal AI agents through your organization's standard OAuth flows, those agents connect with full employee permissions — not narrowly scoped access. BYOA agents connect to work services under individual employee credentials. Services in scope include Google Workspace, Microsoft 365, and Slack. Meta's Muse agent gained connectors for Slack, Linear, and Vercel within weeks of its October 2026 launch. Amazon subsequently blocked Muse for browsing without identifying itself as a bot.
No standard mechanism removes AI agent authorization tokens when an employee departs. An agent authorized by a former employee retains active access until someone manually revokes it.
If agents are connecting to your work systems through employee-authorized OAuth grants, those grants are not automatically revoked when the employee leaves — unless your offboarding procedures explicitly include them.
---