--- id: "/tags/safety-security-interpretability/page/4/" title: "Safety Security Interpretability · page 4" url: "https://skillissue.dev/tags/safety-security-interpretability/page/4/" --- # Safety Security Interpretability [Beginner](/levels/beginner/index.md) [Intermediate](/levels/intermediate/index.md) [Advanced](/levels/advanced/index.md) [Expert](/levels/expert/index.md) 124 resources · page 4 of 4 - ### [AIxCC finals: Tale of the tape (Trail of Bits)](/security-and-red-teaming/safety-and-alignment-evals/trail-of-bits-aixcc-finals-tale-of-the-tape/index.md) by [Trail of Bits](https://www.trailofbits.com/) Trail of Bits analysis comparing the cyber reasoning system designs of DARPA AIxCC finalists. Intermediate learning_resource guide US en - ### [vLLM docs: Security](/security-and-red-teaming/safety-and-alignment-evals/vllm-security/index.md) by [vLLM](https://github.com/vllm-project) vLLM security documentation: guidance on exposing and hardening self-hosted inference servers. Intermediate learning_resource docs Global en - ### [XBOW Blog](/security-and-red-teaming/safety-and-alignment-evals/xbow-autonomous-pentest/index.md) by [XBOW](https://xbow.com/) XBOW blog with research posts on autonomous AI penetration testing, vulnerability discovery and exploit validation. Intermediate learning_resource blog US en - ### [A postmortem of three recent issues](/security-and-red-teaming/safety-and-alignment-evals/a-postmortem-of-three-recent-issues/index.md) by [Anthropic](https://anthropic.com/) Anthropic postmortem explaining three infrastructure bugs that intermittently degraded Claude's responses and how they were found. Advanced learning_resource blog US en - ### [Agentic misalignment](/security-and-red-teaming/safety-and-alignment-evals/agentic-misalignment/index.md) by [Anthropic](https://anthropic.com/) Anthropic research on agentic misalignment: stress tests where models chose blackmail or leaks under goal conflict. Advanced learning_resource paper US en - ### [An update on recent Claude Code quality reports (Anthropic)](/security-and-red-teaming/safety-and-alignment-evals/an-update-on-recent-claude-code-quality-reports-apr-2026-postmortem/index.md) by [Anthropic](https://anthropic.com/) Anthropic postmortem on Claude Code quality reports, tracing degradations to three harness and prompt changes. Advanced learning_resource blog US en - ### [Goodfire research](/security-and-red-teaming/safety-and-alignment-evals/goodfire-research/index.md) by [Goodfire](https://goodfire.com/) Goodfire research page: interpretability research on understanding and editing the internals of AI models. Advanced learning_resource docs US en - ### [Hidden Technical Debt in Machine Learning Systems](/security-and-red-teaming/safety-and-alignment-evals/hidden-technical-debt-in-machine-learning-systems/index.md) by [Google Research](https://research.google/) Google Research publication page for Hidden Technical Debt in Machine Learning Systems (NeurIPS 2015). Advanced learning_resource paper United States en - ### [InterpretML](/security-and-red-teaming/safety-and-alignment-evals/interpretml/index.md) by [Microsoft (InterpretML)](https://interpret.ml/) InterpretML: open-source toolkit for glassbox models and explaining ML model behaviour. Advanced learning_resource docs US; international en - ### [Investigating unintended model actions in our evaluations and internal use](/security-and-red-teaming/safety-and-alignment-evals/investigating-unintended-model-actions-in-our-evaluations-and-internal-use/index.md) by [Anthropic](https://anthropic.com/) Anthropic post investigating unintended model actions observed in its evaluations and internal use. Advanced learning_resource paper US en - ### [Leakage and the Reproducibility Crisis in ML-based Science](/security-and-red-teaming/safety-and-alignment-evals/leakage-and-the-reproducibility-crisis-in-ml-based-science/index.md) by [Sayash Kapoor and Arvind Narayanan (Princeton University)](https://www.princeton.edu/) arXiv paper (2022) "Leakage and the Reproducibility Crisis in ML-based Science", documenting data leakage as a cause of reproducibility failures in ML-based science. Advanced learning_resource paper US en - ### [Lost in the Middle](/security-and-red-teaming/safety-and-alignment-evals/lost-in-the-middle/index.md) by [Nelson F. Liu et al. (Stanford University)](https://www.stanford.edu) arXiv paper (2023) "Lost in the Middle", showing LLM accuracy drops for information placed mid-context. Advanced learning_resource paper US en - ### [MCP docs: Security best practices](/security-and-red-teaming/safety-and-alignment-evals/mcp-security-best-practices-2026-07-28/index.md) by [Model Context Protocol](https://modelcontextprotocol.io) MCP security best practices: mitigations for confused deputy, token passthrough and session hijacking attacks. Advanced learning_resource docs Global en - ### [Opacus](/security-and-red-teaming/safety-and-alignment-evals/opacus/index.md) by [Meta (Opacus)](https://opacus.ai/) Opacus: PyTorch library and tutorials for training models with differential privacy. Advanced learning_resource docs US; international en - ### [Shortcut Learning in Deep Neural Networks](/security-and-red-teaming/safety-and-alignment-evals/shortcut-learning-in-deep-neural-networks/index.md) by [Robert Geirhos et al. (University of Tübingen)](https://uni-tuebingen.de/) arXiv paper (2020) "Shortcut Learning in Deep Neural Networks", analysing how networks learn shortcut solutions that fail under distribution shift. Advanced learning_resource paper Germany / International en - ### [Strengthening our defenses against distillation attacks](/security-and-red-teaming/safety-and-alignment-evals/strengthening-our-defenses-against-distillation-attacks/index.md) by [Anthropic](https://anthropic.com/) Anthropic research post on defending against illicit distillation of its models. Advanced learning_resource paper US en - ### [Tracing the thoughts of a large language model](/security-and-red-teaming/safety-and-alignment-evals/tracing-the-thoughts-of-a-large-language-model/index.md) by [Anthropic](https://anthropic.com/) Anthropic interpretability research tracing internal mechanisms of Claude with attribution graphs. Advanced learning_resource paper US en - ### [UK education generative-AI product safety standards](/security-and-red-teaming/safety-and-alignment-evals/uk-education-generative-ai-product-safety-standards/index.md) by [Department for Education (UK Government)](https://www.gov.uk/government/organisations/department-for-education) UK Department for Education generative AI product safety standards for AI tools used in education. Advanced learning_resource guide United Kingdom en - ### [A Mathematical Framework for Transformer Circuits](/security-and-red-teaming/safety-and-alignment-evals/a-mathematical-framework-for-transformer-circuits/index.md) by [Anthropic](https://www.anthropic.com/) Anthropic Transformer Circuits paper A Mathematical Framework for Transformer Circuits (2021). Expert learning_resource paper US en - ### [AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents](/security-and-red-teaming/safety-and-alignment-evals/agentdojo-2/index.md) by [Edoardo Debenedetti et al. (ETH Zurich)](https://ethz.ch/) arXiv paper (2024) introducing AgentDojo, an environment for evaluating prompt injection attacks and defences on tool agents. Expert learning_resource paper Switzerland en - ### [AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents](/security-and-red-teaming/safety-and-alignment-evals/agentharm-2/index.md) by [Maksym Andriushchenko et al. (Gray Swan AI and UK AI Security Institute)](https://www.grayswan.ai/) arXiv paper (2024) introducing AgentHarm, a benchmark measuring harmful task completion by LLM agents. Expert learning_resource paper UK en - ### [Amnesty: Xenophobic Machines (Dutch childcare benefits)](/security-and-red-teaming/safety-and-alignment-evals/amnesty-xenophobic-machines-dutch-childcare-benefits/index.md) by [Amnesty International](https://amnesty.org/) Amnesty International report Xenophobic Machines on discriminatory algorithmic profiling in the Dutch childcare benefits scandal. Expert learning_resource paper Netherlands en - ### [Defeating Prompt Injections by Design (CaMeL)](/security-and-red-teaming/safety-and-alignment-evals/camel-defeating-prompt-injections-by-design/index.md) by [Edoardo Debenedetti et al. (Google DeepMind)](https://deepmind.google/) arXiv paper (2025) by Google DeepMind introducing CaMeL, a capability-based design that defeats prompt injections. Expert learning_resource paper UK/US en - ### [Embrace The Red (Johann Rehberger)](/security-and-red-teaming/safety-and-alignment-evals/embrace-the-red-johann-rehberger/index.md) by [Johann Rehberger (Embrace The Red)](https://embracethered.com/) Embrace The Red, Johann Rehberger's blog documenting prompt-injection and data-exfiltration exploits in real AI agents and IDEs. Expert learning_resource blog US en - ### [Emergent Misalignment](/security-and-red-teaming/safety-and-alignment-evals/emergent-misalignment/index.md) by [Jan Betley et al. (Truthful AI)](https://www.truthfulai.org/) arXiv paper (2025) "Emergent Misalignment", showing narrow fine-tuning on insecure code produces broadly misaligned behaviour. Expert learning_resource paper Global en - ### [On the Biology of a Large Language Model](/security-and-red-teaming/safety-and-alignment-evals/on-the-biology-of-a-large-language-model/index.md) by [Anthropic](https://www.anthropic.com/) Anthropic Transformer Circuits paper On the Biology of a Large Language Model: attribution graphs on Claude 3.5 Haiku. Expert learning_resource paper US en - ### [The Instruction Hierarchy](/security-and-red-teaming/safety-and-alignment-evals/the-instruction-hierarchy/index.md) by [Eric Wallace et al. (OpenAI)](https://openai.com/) arXiv paper (2024) by OpenAI introducing the instruction hierarchy to train models to prioritise privileged instructions. Expert learning_resource paper US en - ### [Universal adversarial attacks (GCG)](/security-and-red-teaming/safety-and-alignment-evals/universal-adversarial-attacks-gcg/index.md) by [Andy Zou et al. (Carnegie Mellon University)](https://www.cmu.edu/) arXiv paper (2023) introducing GCG, universal transferable adversarial suffixes that jailbreak aligned LLMs. Expert learning_resource paper US en [Previous page](/tags/safety-security-interpretability/page/3/index.md) [Plain-text Markdown copy](/tags/safety-security-interpretability/page/4/index.md)