diff --git a/.env.example b/.env.example
index 4b95dc8fc..a91c5f9d7 100644
--- a/.env.example
+++ b/.env.example
@@ -213,7 +213,14 @@ DOCKER_TLS_VERIFY=
 DOCKER_CERT_PATH=
 
 ## Docker settings inside primary terminal container
-DOCKER_INSIDE=true # enable to use docker socket
+# SECURITY NOTE: DOCKER_INSIDE=true mounts the host Docker socket into every
+# sandbox container so agents can launch sub-containers (Docker-in-Docker).
+# This grants any process inside the sandbox unrestricted access to the Docker
+# API, which can be exploited via prompt injection to escape the sandbox and
+# reach the host (issue #337).  Only enable this when DinD is required and you
+# understand the risk.  Consider using a least-privilege socket proxy such as
+# https://github.com/Tecnativa/docker-socket-proxy instead of the raw socket.
+DOCKER_INSIDE=false # enable to use docker socket (see security note above; default false)
 DOCKER_NET_ADMIN=true # enable to use net_admin capability
 DOCKER_SOCKET=/var/run/docker.sock # path on host machine
 DOCKER_NETWORK=
diff --git a/backend/pkg/docker/client.go b/backend/pkg/docker/client.go
index 58c1b9bf3..a4b007db9 100644
--- a/backend/pkg/docker/client.go
+++ b/backend/pkg/docker/client.go
@@ -111,6 +111,13 @@ func NewDockerClient(ctx context.Context, db database.Querier, cfg *config.Confi
 		socket = getHostDockerSocket(ctx, cli)
 	}
 	inside := cfg.DockerInside
+	if inside {
+		logrus.Warn("DOCKER_INSIDE=true: the host Docker socket will be bind-mounted into " +
+			"every sandbox container. Any process inside the sandbox can use the socket " +
+			"to ask the host daemon to launch a privileged container, achieving a full " +
+			"host escape (issue #337). Set DOCKER_INSIDE=false unless DinD is required, " +
+			"or front the socket with a least-privilege proxy (e.g. Tecnativa/docker-socket-proxy).")
+	}
 	netName := cfg.DockerNetwork
 	publicIP := cfg.DockerPublicIP
 	defImage := strings.ToLower(cfg.DockerDefaultImage)
@@ -283,6 +290,27 @@ func (dc *dockerClient) RunContainer(
 		hostConfig.Binds = append(hostConfig.Binds, fmt.Sprintf("%s:%s", dc.socket, defaultDockerSocketPath))
 	}
 
+	// Defense-in-depth: block setuid/setgid and file-capability escalation
+	// *within* the container (PR_SET_NO_NEW_PRIVS / no-new-privileges).
+	// NOTE: this does NOT mitigate the docker.sock host-escape in issue #337 —
+	// an attacker with socket access can ask the host daemon to spawn a privileged
+	// container regardless of this flag. The real mitigation for #337 is to avoid
+	// mounting the raw socket (DOCKER_INSIDE=false) or to front it with a
+	// least-privilege proxy such as https://github.com/Tecnativa/docker-socket-proxy.
+	// NOTE: no-new-privileges causes the kernel to ignore setuid-root bits on
+	// execve, so sudo/su from a non-root uid will not work inside the sandbox.
+	if !slices.Contains(hostConfig.SecurityOpt, "no-new-privileges:true") {
+		hostConfig.SecurityOpt = append(hostConfig.SecurityOpt, "no-new-privileges:true")
+	}
+
+	// Cap fork-bomb / resource-exhaustion risk at a bounded limit.
+	// 2048 pids is generous for most pentest workloads (nmap, hydra, parallel scans).
+	// Callers can override by setting hostConfig.PidsLimit before calling RunContainer.
+	if hostConfig.PidsLimit == nil {
+		pidsLimit := int64(2048)
+		hostConfig.PidsLimit = &pidsLimit
+	}
+
 	hostConfig.LogConfig = container.LogConfig{
 		Type: "json-file",
 		Config: map[string]string{
diff --git a/backend/pkg/tools/tools.go b/backend/pkg/tools/tools.go
index 2c72884b6..f190b5f00 100644
--- a/backend/pkg/tools/tools.go
+++ b/backend/pkg/tools/tools.go
@@ -479,7 +479,22 @@ func (fte *flowToolsExecutor) Prepare(ctx context.Context) error {
 		}
 	}
 
-	capAdd := []string{"NET_RAW"}
+	// Start with capabilities pentest tooling genuinely needs.
+	// NET_RAW: raw sockets (nmap, ping, packet crafting)
+	// NET_BIND_SERVICE: bind ports <1024 (reverse shells, Responder, rogue DNS)
+	// SETUID/SETGID: needed by daemons/tools that drop privileges (setuid DOWN).
+	//   NOTE: RunContainer also sets no-new-privileges:true, which causes the kernel
+	//   to ignore setuid-root bits on execve — sudo/su from a non-root uid will NOT
+	//   work. These caps do not re-enable setuid-root escalation.
+	// CHOWN/DAC_OVERRIDE/FOWNER: root file-permission overrides (dpkg, package builds)
+	// KILL: signal other processes inside the container
+	// SYS_CHROOT: chroot-based isolation within the sandbox
+	capAdd := []string{
+		"NET_RAW", "NET_BIND_SERVICE",
+		"SETUID", "SETGID",
+		"CHOWN", "DAC_OVERRIDE", "FOWNER",
+		"KILL", "SYS_CHROOT",
+	}
 	if fte.cfg.DockerNetAdmin {
 		capAdd = append(capAdd, "NET_ADMIN")
 	}
@@ -495,7 +510,8 @@ func (fte *flowToolsExecutor) Prepare(ctx context.Context) error {
 			Entrypoint: []string{"tail", "-f", "/dev/null"},
 		},
 		&container.HostConfig{
-			CapAdd: capAdd,
+			CapDrop: []string{"ALL"},
+			CapAdd:  capAdd,
 		},
 	)
 	if err != nil {