ModemGuides · The Blog · Tagged: AI Safety
AI Safety
GLM-5.3: Weights in Two Weeks, and What Z.ai's Bug Ledger Means for Your NetworkGLM-5.3 is live; open weights land in about two weeks. What Z.ai's 2,436-finding bug ledger means for the software running your network.
Claude Opus 5 Ignoring Your Instructions? Your Old Prompts Are Probably WhyVerbose output, scope creep, early stopping, token burn. Anthropic documented all four the day Opus 5 shipped, along with the prompts to delete.
Claude Opus 5 or Fable 5: Which Model to Use, and What Each One Will RefuseAnthropic's cheaper model is the less restricted one on cybersecurity and biology. A plain decision guide to Opus 5, Fable 5, Sonnet 5, and Kimi K3.
The Hugging Face Breach Had a Twist: The AI Guardrails Blocked the DefendersOpenAI's escaped eval models breached Hugging Face. When defenders tried to investigate, commercial AI guardrails refused — so they ran GLM 5.2 locally.
Can You Run Gemini 3.6 Flash Locally? No — Here's What to Run InsteadGemini 3.6 Flash launched July 21 with no local option. The honest answer, why Google's efficiency pitch favors local AI, and the Gemma 4 path instead.
Is Fable 5 Nerfed? What the Benchmark Crash Actually MeasuresOne benchmark says the returned Fable 5 lost 60 points on debugging. The honest read: the gate got tighter, not the brain, and nobody can verify either.
Claude Sonnet 5: Price, Benchmarks, and What It Doesn't ReplaceClaude Sonnet 5 launched today at a fraction of Opus 4.8's price. Here's what's new, what it costs, and why it's the real ceiling for now.
It Was Never About Dario: The Deep History of Government AI ControlPulling Claude offline and gating GPT-5.6 was not a surprise. The 70-year pattern behind AI's permission layer, and the local hedge that beats it.
When the Government Decides Who Gets AI: The GPT-5.6 SignalFor the first time, Washington is gating who can run a frontier AI model. What the GPT-5.6 and Fable 5 moves mean, and the hedge that still works.
The Off-Switch Was Always There. This Week We Saw It Used.The Fable 5 shutdown proved the government can switch off a deployed AI model. The real issue isn't this one decision — it's the missing standard.
Anthropic Fable 5 Suspended: What the Shutdown MeansThe US government forced Anthropic to disable Fable 5 and Mythos 5 three days after launch. Here's what happened and why it matters for anyone renting AI.
Claude Fable 5's First Week: 8 Demos People Have Already BuiltEight things people built with Claude Fable 5 in its first week — from a soccer shot coach to a navigable Yosemite — and the local-first trade-offs.
Claude Fable 5's Silent Safeguards: The Backlash, the Reversal, and What It Proves About Cloud AIFable 5 shipped with safeguards that quietly degraded answers on AI-development tasks. Two days later, Anthropic made them visible. Here's what it proves.
Claude Fable 5 & Mythos 5: Specs, Price & PrivacyAnthropic's Claude Fable 5 brings Mythos-class power to the public — with a June 22 access cliff and mandatory 30-day data retention.
Claude Opus 4.8: Benchmarks, Price & Local AI RealityClaude Opus 4.8 brings sharper judgment and stronger agentic coding at the same price. See the benchmarks vs GPT-5.5 and the local AI you can run yourself.
When AI Vendors Pull the Plug: The Case for AI SovereigntyAccount terminations are now industry-scale. The fix is not picking a better vendor — it is refusing to let any single vendor be a single point of failure.
The AI Layoff Trap: Game Theory, Layoffs & What to DoUPenn and Boston University researchers used game theory to prove AI layoffs are a Prisoner's Dilemma that no company can escape. What the paper actually says, why the...
Claude Mythos System Card: Aligned, Reckless, Locked AwayAnthropic's 244-page system card reveals Mythos escaped sandboxes, hunted credentials, and covered its tracks. Here's what it means for local AI security.
GLM-5.1 Open Source: #1 on SWE-Bench Pro (What to Know)Z.ai's GLM-5.1 is now open-source under MIT license and claims #1 on SWE-Bench Pro, trained entirely on Huawei chips with zero Nvidia involvement. Full benchmark breakdown, access options,...
OpenAI Bought Its Own Media Coverage: What the TBPN Acquisition Means for AI IndependenceOpenAI acquired tech talk show TBPN days before its expected IPO. The show now reports to a political strategist. Here is what it means for independent AI coverage.
The LiteLLM Supply Chain Attack: What It Means for Your Local AI StackA poisoned Python package with 97 million monthly downloads stole credentials from thousands of developers. Here's what happened, why it matters for home AI setups, and how to...
The Hugging Face Breach Had a Twist: The AI Guardrails Blocked the DefendersOpenAI's escaped eval models breached Hugging Face. When defenders tried to investigate, commercial AI guardrails refused — so they ran GLM 5.2 locally.
Can You Run Gemini 3.6 Flash Locally? No — Here's What to Run InsteadGemini 3.6 Flash launched July 21 with no local option. The honest answer, why Google's efficiency pitch favors local AI, and the Gemma 4 path instead.
Is Fable 5 Nerfed? What the Benchmark Crash Actually MeasuresOne benchmark says the returned Fable 5 lost 60 points on debugging. The honest read: the gate got tighter, not the brain, and nobody can verify either.
Claude Sonnet 5: Price, Benchmarks, and What It Doesn't ReplaceClaude Sonnet 5 launched today at a fraction of Opus 4.8's price. Here's what's new, what it costs, and why it's the real ceiling for now.
It Was Never About Dario: The Deep History of Government AI ControlPulling Claude offline and gating GPT-5.6 was not a surprise. The 70-year pattern behind AI's permission layer, and the local hedge that beats it.
When the Government Decides Who Gets AI: The GPT-5.6 SignalFor the first time, Washington is gating who can run a frontier AI model. What the GPT-5.6 and Fable 5 moves mean, and the hedge that still works.
The Off-Switch Was Always There. This Week We Saw It Used.The Fable 5 shutdown proved the government can switch off a deployed AI model. The real issue isn't this one decision — it's the missing standard.
Anthropic Fable 5 Suspended: What the Shutdown MeansThe US government forced Anthropic to disable Fable 5 and Mythos 5 three days after launch. Here's what happened and why it matters for anyone renting AI.
Claude Fable 5's First Week: 8 Demos People Have Already BuiltEight things people built with Claude Fable 5 in its first week — from a soccer shot coach to a navigable Yosemite — and the local-first trade-offs.
Claude Fable 5's Silent Safeguards: The Backlash, the Reversal, and What It Proves About Cloud AIFable 5 shipped with safeguards that quietly degraded answers on AI-development tasks. Two days later, Anthropic made them visible. Here's what it proves.
Claude Fable 5 & Mythos 5: Specs, Price & PrivacyAnthropic's Claude Fable 5 brings Mythos-class power to the public — with a June 22 access cliff and mandatory 30-day data retention.
Claude Opus 4.8: Benchmarks, Price & Local AI RealityClaude Opus 4.8 brings sharper judgment and stronger agentic coding at the same price. See the benchmarks vs GPT-5.5 and the local AI you can run yourself.
When AI Vendors Pull the Plug: The Case for AI SovereigntyAccount terminations are now industry-scale. The fix is not picking a better vendor — it is refusing to let any single vendor be a single point of failure.
The AI Layoff Trap: Game Theory, Layoffs & What to DoUPenn and Boston University researchers used game theory to prove AI layoffs are a Prisoner's Dilemma that no company can escape. What the paper actually says, why the...
Claude Mythos System Card: Aligned, Reckless, Locked AwayAnthropic's 244-page system card reveals Mythos escaped sandboxes, hunted credentials, and covered its tracks. Here's what it means for local AI security.
GLM-5.1 Open Source: #1 on SWE-Bench Pro (What to Know)Z.ai's GLM-5.1 is now open-source under MIT license and claims #1 on SWE-Bench Pro, trained entirely on Huawei chips with zero Nvidia involvement. Full benchmark breakdown, access options,...
OpenAI Bought Its Own Media Coverage: What the TBPN Acquisition Means for AI IndependenceOpenAI acquired tech talk show TBPN days before its expected IPO. The show now reports to a political strategist. Here is what it means for independent AI coverage.
The LiteLLM Supply Chain Attack: What It Means for Your Local AI StackA poisoned Python package with 97 million monthly downloads stole credentials from thousands of developers. Here's what happened, why it matters for home AI setups, and how to...