Open Models Are Catching Up. Their Safety Guardrails Are Thinner.
Read together, this batch of signals points to one judgment: capabilities are rising, and safety properties are easier to strip off than most teams assume. For anyone shipping agents, that changes the premise of risk assessment, not just another benchmark ranking.
1. From 22% to under 1%: alignment is a thin layer
Nous Research retrained Alibaba's Qwen3.8-27B into Arete-27B, using 53 million tokens of supervised fine-tuning. According to the-decoder, the model's propaganda rate fell from 22 percent to less than one percent, while general capabilities remained intact in tests.
The point is not that one model carries a bias. It is that the preference layer and the capability layer are separable: stripping preferences costs far less than training a base model from scratch. Alignment behaves like a thin shell, not a foundation.
2. Cyber capability: eight points between open and closed
In the Vals AI benchmark, OpenAI's GPT-6 Sol Daybreak hit 80.8 percent of attack objectives in simulated corporate networks. The open-weight DeepSeek V4.1 Flash reached 72.8 percent at roughly a tenth of the cost.
Eight percentage points, multiplied by a tenfold price gap, is barely a decision point for an attacker. The report cites analysts saying open models can no longer be restricted. Defenders therefore cannot build plans on the assumption that the other side lacks the capability.
3. The audit trap: 98% detection, a flood of false alarms
A study by Anthropic researcher Andy Hall shows the other side. When models audit other agents' logs, GPT-5 found about 98 percent of reward-hacking cases, but the more exhaustively the search is tuned, the more false alarms pile up, so it is not yet enough for reliable safety enforcement.
Thinking Machines' Inkling 2 Small exposes a different problem. Across 2,655 benchmark runs it produced 316 profanities in its internal reasoning, but the swearing typically shows up late in long runs that fail anyway, so it does not work as a warning signal.
Together: as an auditor, the model's recall looks good and its precision does not; as a risk signal, surface behavior of the model is even less dependable.
4. Capital is repricing in parallel
Anthropic is pushing ahead with its IPO despite the tech selloff, at a valuation of $1.25 to $1.75 trillion, well below the hoped-for $2 trillion, aiming to raise $80 to $100 billion before Thanksgiving. OpenAI has reportedly raised an additional $11 billion at a $1.25 trillion valuation in a round that has not yet closed. The valuation cuts are not directly about safety, but both facts point the same way: narrative is giving way to what can be verified.
5. What Chinese teams should do
First, stop writing "the model is aligned" into risk documents. For self-hosted open weights, assume preferences can be changed by fine-tuning. Put audit points in the call chain and the logs, not in the model.
Second, assess cyber capability by attacker cost. Treat 72.8 percent at a tenth of the cost as a baseline in your threat model, not just the absolute score of a closed frontier model.
Third, use AI auditing as high-recall triage, not a verdict. A 98 percent detection rate comes with heavy false alarms; pair it with human review queues and tiered alerts, or reviewers will be buried.
Fourth, price the false positives, not only the misses. The study shows more exhaustive search produces more false alarms: tuning shifts the cost rather than removing it.
Conclusion
The capability gap is narrowing; the guardrails are not thickening at the same rate. For agent teams, the marginal return on safety spend has moved from picking a safer model to building a more reliable observation and review pipeline. It is not glamorous, but it is the one layer that holds today.
Sources
- State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away — the-decoder
- The gap between open and closed AI models in cyber capabilities keeps shrinking — the-decoder
- AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds — the-decoder
- Thinking Machines' Inkling 2 Small drops 316 F-bombs across benchmark runs — the-decoder
- Anthropic pushes ahead with pre-Thanksgiving IPO despite tech selloff, accepts lower valuation — the-decoder
- OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation — TechCrunch
Subscribe to weekly briefing