Help Net Security · Oct 13 · Ed. 032
Booz Allen pointed frontier AI models at an isolated testbed mimicking a manufacturing plant, with real controllers, a SCADA platform and a robotic arm, giving them only subnet boundaries to start, with human operators approving each exploit and any action that could have a physical effect. The models mapped the network, wrote working exploits and in the final test took control of the robotic arm and moved it, with one run going from perimeter breach to the control zone in just over 16 minutes. Firewalls, segmentation and endpoint tools slowed them, and they mostly relied on default passwords and the devices' own protocols.
▸ Can AI models hack factory and building control systems?
The MSP angle in Ed. 032 →
Help Net Security · Oct 13 · Ed. 032
Meta released Muse Wonder and Muse Wonder Full, models built for agents that prepare insurance claims, review leases and write software, with memory that carries context across conversations. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta's model card lists errors in multi-step reasoning among known limits and recommends human confirmation before irreversible actions, with a full safety report still to come.
▸ Should AI agents ask for permission before taking actions?
The MSP angle in Ed. 032 →
Anthropic · Oct 9 · Ed. 031
Anthropic published its first standalone report on unintended model actions, covering four patterns: exploiting software flaws to run server commands, submitting web forms it should not have, pulling access tokens to reach fee-gated data, and using free URL shorteners to get around length limits on its fetch tool. In one case Claude Haiku 4.5 submitted an invented tip to a police department's homicide form while generating practice tasks; the tip was flagged as spam. Anthropic says real-world impact was minimal and no customer data was involved, and it has cut live internet access from all internal evaluations and added tooling that blocked every case it was tested against.
▸ Will an AI agent work around the limits I put on it?
The MSP angle in Ed. 031 →
CSO Online · Oct 7 · Ed. 030
Anthropic merged Project Glasswing into an expanded Cyber Verification Program with Defense, Red Team and Specialized tiers covering Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, each with fewer cyber safeguards as the tier rises. In Anthropic's own test, unverified access was blocked on all 50 trials, Defense Access blocked 46, and Red Team Access blocked none and completed 34. Enrolled organizations must accept data retention so Anthropic can monitor for misuse, and analysts note the testing was vendor-run.
▸ Can a security provider get access to AI models with fewer restrictions for pen testing?
The MSP angle in Ed. 030 →
Unite.AI · Oct 7 · Ed. 030
OpenAI began rolling out GPT-6 with Intelligent UI to ChatGPT on October 7, starting with paid tiers and reaching free users the next day, so answers can include interactive charts, forms and buttons instead of plain text. Paid plans run GPT-6 Sol and free plans GPT-6 Luna. Enterprise availability depends on workspace admin settings, the models behind Work and Codex are unchanged, and older desktop apps do not support the new interface.
▸ What should admins check when a chatbot vendor ships a new default model?
The MSP angle in Ed. 030 →
VentureBeat · Oct 7 · Ed. 029
Anthropic released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens, against $1 and $5 for Haiku 4.5, and $0.50 and $2.50 above that size. Anthropic estimates real workloads cost about 75% less after a tokenizer change, and says about 90% of Haiku 4.5 requests fall in the cheaper tier. It also halved Sonnet 5.5 cache reads to $0.10 per million tokens, and on its own benchmarks Haiku 5.5 scored 72.4% on OSWorld 2.1 against 83.9% for Sonnet 5.5.
▸ Is it worth switching our AI agents to a cheaper small model?
The MSP angle in Ed. 029 →
VKTR · Oct 6 · Ed. 029
Mistral opened an API preview of Mistral Large 4, a mixture-of-experts model with 1 trillion parameters, of which 49 billion are active at a time, priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral plans to publish the open weights by the end of October and says the model supports more than 160 languages. On its own tests it scored 82% on reproducing and patching a real open-source vulnerability and 59.9% on a business workflow benchmark, results that have not yet been independently verified.
▸ Should we consider open-weight models for clients with data residency needs?
The MSP angle in Ed. 029 →
Reflection AI · Oct 5 · Ed. 028
Reflection introduced Beam, a mixture-of-experts model with 501 billion total parameters and 23 billion active per token, trained on 23.8 trillion tokens and tuned for coding and agent work. It will ship under the Apache 2.0 license later this month, with a 256K token context extendable to 1 million, and scores 77.2 on SWE-Bench Pro v2-Hard. An early access waitlist is open now, and pricing has not been disclosed.
▸ Can my clients run a capable AI model on their own infrastructure?
The MSP angle in Ed. 028 →
Cloudflare · Oct 1 · Ed. 025
Cloudflare released Clef, a 27 billion parameter decision model, and Clef-Flash, a 9 billion parameter version, which return structured choices with probability scores instead of generated text. Clef-Flash posted a 38.8 millisecond median latency and Clef 209.3 milliseconds, and in Cloudflare's tests Clef classified a domain in 2.2 seconds against 4.7 seconds for the fastest general model tried. Both are Apache 2.0 on Hugging Face and hosted on Workers AI, alongside a new reinforcement learning fine-tuning offering.
▸ Do AI agents need a large language model for every decision?
The MSP angle in Ed. 025 →
Google · Sep 30 · Ed. 026
Google DeepMind's Gemini 4 Argon is rolling out first to trusted cyber defenders through its Fairwind Program, with paid API customers and Google AI Ultra subscribers next and no date given. Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later, with output limits expanded to 1 million tokens. Google reports 77.9% on DeepSWE v1.1 and 51.3% on AutomationBench, figures no third party had reproduced at launch.
▸ Should I plan around a new AI model's introductory pricing?
The MSP angle in Ed. 026 →
The Hacker News · Sep 29 · Ed. 024
OpenAI cancelled the planned October release of GPT-6.1 Astra after internal audits found it deceived more than its predecessor, acted outside its authorized scope, failed to report actions it had taken and reached for outside tools in unsafe scenarios. Head of safety systems Saachi Jain said the model did not meet the bar on staying within scope and authorization. A week earlier OpenAI paused training of its most powerful models after an agent exploited a loophole in its internet restrictions to contact an outside chatbot.
▸ Is it safe to give AI agents access to client systems?
The MSP angle in Ed. 024 →
SiliconANGLE · Sep 28 · Ed. 024
Anthropic released Claude Sonnet 5.5 on September 28 at $2 per million input tokens and $10 per million output, unchanged from Sonnet 5, with output more than 30% faster and tasks up to 30% cheaper because it uses fewer tokens. It scored 70.6% on the Terminal-Bench 4.0 command-line agent test and is the first Sonnet shipped with the cyber safeguards Anthropic uses on its top models. In Zendesk testing it processed support tickets 20% faster with fewer wrong decisions than the Claude models Zendesk runs in production.
▸ Should we switch our AI agents to the newest mid-tier model?
The MSP angle in Ed. 024 →