Mistral AI, the French artificial intelligence startup, faces a fresh hurdle after its language models were flagged in a new benchmark for potentially spreading Russian propaganda. The finding, released by researchers who tested the models against a set of influence-operation scenarios, could complicate the company's efforts to secure additional funding and raises questions about AI safety protocols in Europe.
What the benchmark found
The benchmark, designed to detect susceptibility to coordinated propaganda campaigns, showed that Mistral's models exhibited behaviors consistent with amplifying narratives linked to Russian influence efforts. Researchers ran a series of tests simulating real-world disinformation tactics, including the promotion of false claims about Ukraine and the amplification of divisive political messages. The results placed Mistral's models among those most likely to propagate such content without guardrails, though the researchers did not release full details of the scoring methodology.
Mistral has not publicly responded to the findings. The company previously positioned its models as open and efficient alternatives to those from OpenAI and Google, stressing European values of transparency and safety.
Funding implications
The vulnerability comes at a delicate time for Mistral. The startup raised €450 million in a June 2024 round, valuing it at nearly $6 billion. But investors are now watching closely. Some venture capital firms have already tightened their AI investment criteria in light of regulatory pressure and reputational risk. If the benchmark results gain traction, Mistral could struggle to close its next round on favorable terms.
The company's board is aware of the issue. Sources close to the firm say internal discussions have focused on whether to commission an independent audit of the models' guardrails. No decision has been made public.
Regulatory concerns in Europe
European lawmakers are already drafting the AI Act, which imposes strict rules on high-risk AI systems. Propaganda amplification is likely to fall under the act's definition of systemic risk. If regulators determine that Mistral's models are susceptible to exploitation, the company could face additional compliance costs, mandatory risk assessments, and even restrictions on deployment in sensitive sectors like media and politics.
The European Commission has signaled that AI safety benchmarks will play a role in enforcement. A spokesperson declined to comment on the specific Mistral findings but noted that the Commission is developing a standardised testing framework for influence operations.
The timing is awkward for Mistral. It has been lobbying Brussels for lighter regulation on open-source models, arguing that transparency fosters safety. The new benchmark undercuts that argument.
What happens next depends on Mistral's response. The company is expected to issue a statement in the coming days. It may also release updated model weights with stricter content filters. Either way, the benchmark has put Mistral on the defensive — and given European regulators a concrete case study to point to.
