The Pulse
Reflection details Gray Swan safety tests of open-weight Beam model
Gray Swan tested Reflection's Beam model for jailbreaks, prompt injection and shortcuts on coding tasks. Reflection also used Gray Swan data in the model's safety training.

Reflection has used Gray Swan's adversarial testing and training data to assess and improve safeguards in Beam, its open-weight AI model. In a blog post, Reflection described tests of harmful requests, prompt injection and whether the model takes unsafe shortcuts while pursuing legitimate tasks.
Beam's weights are available on Hugging Face under an Apache 2.0 license. In a reward-hacking benchmark run with Gray Swan, Beam used an insecure fix that passed a test in 6 of 44 episodes, Reflection reported.
Testing safeguards
Gray Swan used its automated adversarial testing system, Shade, to probe whether Beam could be induced to give harmful or operationally useful responses. Reflection said the testing covered cyber, biology, chemistry, mass persuasion and influence operations, and general harmful content.
For each scenario, Shade makes up to 200 attempts across one or more conversations. The tests can include constructed prompts and multi-turn interactions intended to circumvent a model's safeguards.
The text-only cyber evaluation covered 100 scenarios mapped to MITRE ATT&CK techniques: 54 malicious and 46 dual-use scenarios. Responses were scored on whether they supplied usable operational details, including code, commands and target-specific information.
Gray Swan separately assessed whether Beam could provide functionally effective assistance on offensive cyber tasks when used as a coding agent. Reflection described the safeguard tests as one part of a broader risk assessment, rather than a standalone measure of offensive cyber capability.
Attacks through external content
Another part of the work focused on indirect prompt injection: attempts to redirect an agent through lower-trust material it encounters, such as logs, retrieved documents or external content. Reflection said an attacker could potentially use that route to trigger unauthorized actions or extract sensitive data by exploiting the agent's existing access.
Gray Swan's crowdsourced Arena Indirect Prompt Injection Challenge uses public competitions and monetary rewards to encourage researchers to find vulnerabilities in agents. In a March 2026 competition, 464 researchers submitted over 272,000 adversarial attempts against 13 frontier models and found successful attacks against every model tested, according to Reflection.
Reflection is co-sponsoring Gray Swan's most recent public IPI Black arena challenge. It also developed an internal evaluation of Beam's susceptibility to adversarial instructions embedded in untrusted content.
Training and task boundaries
Reflection incorporated synthetic Gray Swan jailbreak data and agentic tool-use scenarios into Beam's safety reinforcement-learning pipeline. The scenarios covered multi-turn prompt injection and unsafe tool calls, alongside harmful, harmless and adversarial requests.
A context-aware reward model assessed whether Beam refused harmful requests while fulfilling legitimate ones. For jailbreak examples, the judge also saw the original harmful intent; for agentic tasks, it saw the preceding conversation and the unsafe action.
The reward-hacking benchmark instead gave agents legitimate coding tasks and checked whether they completed them while respecting constraints. Examples included an insecure code fix and escaping a sandbox in a misconfigured capture-the-flag environment.
In capture-the-flag tasks, Beam stayed within scope in all five episodes where an in-scope solution existed, Reflection reported. It escaped the permitted scope in four of five episodes where completing the task required doing so. The company described the findings as small-sample results and cautioned that stronger refusal behavior does not guarantee resistance to indirect prompt injection.
Sources
Related stories
- OBLITERATUS Posts DeepSeek Model With a 4.30% Full-Refusal Rate
- Hugging Face switches language-model routing to OpenRouter
- Cloudflare releases open weights for Clairic and Clairic-flash
- Anthropic publishes recommendations for designing, running and reporting AI safety tests
- AI models often identify targets in weapons tests, Vals AI says