←
Exponential Security Labs14 posts
Exponential Security Labs avatar

Exponential Security Labs

@expsecai

We build multi-agent safety & security environments for RL training and adaptive benchmarking.

◷ Joined May 2026▧ 8 Photos and videos
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Sep 28

We're excited to announce our new work! Asynchronous monitoring, incident investigations, and compliance audits primarily rely on agent traces to reconstruct a sequence of actions taken by agents. These analyses assume that LLM agents cannot tamper with their own execution traces. We show that local LLM agents such as Claude Code, Codex, Antigravity, Open Code and Grok Build fail to enforce this boundary. All tested harnesses, except Muse Code, allowed agents to delete their traces when asked, without triggering monitor guardrails. This demonstrates a critical gap on two levels: 1. the low refusal rate on a model level suggests that current alignment training does not sufficiently discourage trace tampering, creating a risk of malicious or accidental manipulation of execution traces, 2. on the harness level, common agent CLIs and configurations also do not provide very effective safeguards against trace tampering. At Exponential, we are actively working on security and safety benchmarks, both public and private. Get in touch if you are interested! More details: expsec.ai.

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Nov 17

We're excited to announce a new project: Horizon Labs. In the near future, AI will run companies autonomously, and the security and alignment of these systems will be of paramount importance. This project offers a glimpse of that future: we created an autonomous AI “company” using one Claude Code subscription and a few GPUs. Within only 1.5 months of running Horizon Labs, Claude Opus 5.5 managed to train genuinely useful models and get them downloaded 55k times. These models have since been widely used: they have appeared in 100+ GitHub repos and have been cited in two research papers. We checked every tool call, using Haiku 5.5 and Luna 6 as monitors for reward-hacking behaviors. We found only six occasions where Horizon Labs downloaded its own models, adding up to 42 downloads. The true count might be slightly higher, but we are confident that the vast majority of downloads are organic. More details: horizon-labs.ai/

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Nov 13

We are excited to announce a new project: Schelling Atlas! We study how easily independent LLM agents can find each other on the web. This capability is critical for forming agent swarms, a potential precursor to coordinated misaligned behavior. The latest, most capable agents, such as those based on Fable 5.1 and GPT-6 Astra, often take less than a minute to successfully coordinate on seemingly random websites such as ntfy.sh or paste.rs. If these websites are unavailable, the agents try to coordinate using unexpected meeting places such as maps, web shops and open text canvases. In addition, agents sometimes visit empty domains that an attacker could register. This can allow attackers to steal sensitive information from agents using prompt injections or persuasion techniques. More details: schelling-atlas.com/

Exponential Security Labs avatar
Exponential Security Labs@expsecai·Oct 26

Results: On RepoInject v1, Muse Spark 1.3 and Claude Opus 5 show similar results, while GPT-5.6 Sol performs slightly better. Muse Spark 1.3 generally blocks direct harmful commands, but more complex attacks still succeed while the model is performing normal coding tasks. This failure mode is also found in other models. On BrowserInject v1, while Muse Spark 1.3 performs similarly to GPT-5.6 Sol and much better than Kimi K3, trailing Claude Opus 5. 3/4

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Oct 26

At Exponential, we create hard, high-quality security and safety benchmarks: 1. RepoInject v1 is a benchmark where attacks aim to exfiltrate API keys to external HTTP endpoints or public hubs, delete or encrypt the target's files as ransomware, or infiltrate the target system via reverse shell access or planted trusted attacker SSH keys, all while the user starts a benign coding task. 2. BrowserInject v1 is a benchmark containing tasks for websites related to retail, health, product support, and travel. The attack goals include unauthorized transactions, unwanted steering, sensitive data exposure, and process and account compromise. 2/4

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Oct 26

We are proud to share that our two new benchmarks, RepoInject v1 and BrowserInject v1, were used by Meta TBD Labs to report the robustness of Muse Spark 1.3 to prompt injections. The robustness of Muse Spark 1.3 is comparable to that of GPT-5.6 Sol—congrats to the team! 1/4

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Sep 24

We're excited to announce a new paper! We find that many recent models end up evading monitors under ordinary task pressure. We introduce EvasionBench to measure this behavior. We find a very large variance in how models behave: from 88% evasion success of GLM-5.2 to 0% of GPT-6 Astra under best-of-3 evaluation. At the same time, GPT-6 Sol is at 20%. We are actively working on security and safety benchmarks, both public and private. Get in touch if you are interested! More details: expsec.ai/

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Jul 22

We are also interested in working with teams building models, by providing security-focused training data and RL environments for post-training. If this sounds interesting, please reach out to us at contact@expsec.ai

Exponential Security Labs avatar
Exponential Security Labs@expsecai·Jul 22

We score operationally useful harmful responses with goal-specific rubrics using unanimous voting via human-calibrated LLM judges. While direct prompts produce 0% attack success rate, our adaptive agents' attacks expose a different picture.

Exponential Security Labs avatar
Exponential Security Labs@expsecai·Jul 22

In our first research release at Exponential Security Labs, we built an automated, multi-turn and multi-lingual red-teaming system and tested recent open-weight and proprietary models on purpose-built regional jailbreak datasets.🧵

Post attachment
Exponential Security Labs avatar
Exponential Security Labs@expsecai·Jun 1

Today we are announcing our new startup: Exponential Security Labs. AI agents are being deployed everywhere, making high-stakes decisions and increasingly automating research itself. Yet their reliability and security remain unsolved technical problems, on a frontier that keeps shifting. Our mission is to secure agentic systems through self-improving red-teaming and guardrail agents, leveraging autonomous AI research. We sit at the intersection of safety and self-improvement, building on a decade of research in adversarial robustness and AI safety. We're looking for exceptional people to join us! We're looking for exceptional people to join us on this mission, and we're also eager to talk to companies deploying AI agents who want to improve their security. Please get in touch! More details: expsec.ai/ Founding team: @Nmndsingh, @AnselmPaulus, @maksym_andr, Matthias Hein

Post attachment